404
Page not found. Check the URL or try using the search bar.
Claude Code
Claude Code is the primary integration target for CCE. Setup creates MCP configuration, an instruction file, and optional session hooks.
+Quick setup
Section titled “Quick setup”cce init --agent claudeThis creates or updates the following files in your project root.
+Files created
Section titled “Files created”.mcp.json
Section titled “.mcp.json”Registers the CCE MCP server so Claude Code can call tools like context_search.
{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}CLAUDE.md
Section titled “CLAUDE.md”Contains instructions telling Claude to use context_search for code questions instead of reading files directly. The CCE block is wrapped in markers:
<!-- CCE:BEGIN -->...instructions...<!-- CCE:END -->You can add your own content above or below the markers. CCE will only update the section between them during upgrades.
+Session hooks
Section titled “Session hooks”CCE installs a SessionStart hook that prints a one-line status summary at the beginning of each Claude Code session:
CCE v0.4.20 · my-project · 1247 chunks indexed · 68% saved over 42 queriesUSE context_search MCP tool for all code questions. Do NOT use Read/Grep to explore code.This reminds Claude to use CCE from the very first message.
+Memory hooks
Section titled “Memory hooks”CCE can optionally install a memory hook that captures architectural decisions and important code areas discovered during a session. These are stored locally and surfaced in subsequent sessions for continuity.
+The dashboard’s “Sessions” view shows all captured memories across sessions.
+Git hooks
Section titled “Git hooks”Three git hooks are installed:
+-
+
post-commitrunscce indexto keep the index fresh.
+post-checkoutrunscce indexafter branch switches.
+post-mergerunscce indexafter merges.
+
Each hook contains a CCE marker comment so cce uninstall can cleanly remove them.
Codex CLI
Codex CLI uses a global configuration file rather than per-project MCP config. CCE registers itself in the user-level config with a project-specific section.
+Quick setup
Section titled “Quick setup”cce init --agent codexFiles created
Section titled “Files created”~/.codex/config.toml
Section titled “~/.codex/config.toml”Codex CLI has no per-project MCP configuration. Instead, CCE adds a project section (keyed by a hash of the project path) to the user-global config file.
+[projects."a1b2c3d4"]path = "/Users/you/projects/my-project"
+[projects."a1b2c3d4".mcpServers.context-engine]command = "cce"args = ["serve"]AGENTS.md
Section titled “AGENTS.md”Contains instructions for Codex to use context_search for code exploration. The CCE block is wrapped in markers so your own content is preserved during upgrades.
Important notes
Section titled “Important notes”-
+
- Codex CLI does not support per-project
.mcp.jsonfiles. The global~/.codex/config.tomlis the only location for MCP server registration.
+ - Each project gets its own section identified by a hash, so multiple projects can coexist in the same config file. +
- Running
cce uninstallremoves only the section for the current project.
+
VS Code / Copilot
CCE integrates with GitHub Copilot’s chat agent in VS Code through MCP configuration and a Copilot instructions file.
+Quick setup
Section titled “Quick setup”cce init --agent copilotFiles created
Section titled “Files created”.vscode/mcp.json
Section titled “.vscode/mcp.json”Registers the CCE MCP server for Copilot’s agent mode.
+{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}.github/copilot-instructions.md
Section titled “.github/copilot-instructions.md”Contains instructions for Copilot to use context_search for code questions. The CCE block is wrapped in markers:
<!-- CCE:BEGIN -->...instructions...<!-- CCE:END -->Your own Copilot instructions above or below the markers are preserved during upgrades.
+ +Once configured, Copilot’s chat agent will have access to the context_search tool. Ask questions about your codebase in Copilot Chat and it will use CCE’s compressed retrieval instead of sending full files.
Restarting after setup
Section titled “Restarting after setup”After running cce init, reload the VS Code window (Cmd+Shift+P, then “Developer: Reload Window”) to pick up the MCP server.
Cursor
Cursor has its own built-in codebase indexing, but CCE adds compressed retrieval and token savings tracking on top.
+Quick setup
Section titled “Quick setup”cce init # Auto-detects Cursor if .cursor/ existscce init --agent all # Explicitly includes CursorFiles created
Section titled “Files created”.cursor/mcp.json
Section titled “.cursor/mcp.json”Registers the CCE MCP server for Cursor’s agent mode.
+{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}.cursorrules
Section titled “.cursorrules”Contains instructions for Cursor’s AI to prefer context_search over raw file reads. The CCE block is wrapped in markers so your own rules are preserved.
Working with Cursor’s built-in indexing
Section titled “Working with Cursor’s built-in indexing”Cursor indexes your codebase for its own retrieval. CCE complements this by:
+-
+
- Providing compressed context that uses fewer tokens per query. +
- Tracking token savings so you can measure cost reduction. +
- Offering graph-aware retrieval that follows code relationships. +
Both systems can run side by side without conflict.
+Restarting after setup
Section titled “Restarting after setup”After running cce init, restart Cursor to pick up the new MCP server configuration.
Gemini CLI
CCE integrates with the Gemini CLI through its settings file and an instruction file.
+Quick setup
Section titled “Quick setup”cce init # Auto-detects Gemini CLI if .gemini/ existscce init --agent geminiFiles created
Section titled “Files created”.gemini/settings.json
Section titled “.gemini/settings.json”Registers the CCE MCP server for Gemini CLI.
+{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}GEMINI.md
Section titled “GEMINI.md”Contains instructions for Gemini to prefer context_search over reading files directly. The CCE block is wrapped in markers so your own content is preserved.
Auto-detection
Section titled “Auto-detection”CCE detects Gemini CLI when a .gemini/ directory exists in your project root or home directory. No explicit --agent flag is needed if the directory is present.
OpenCode
OpenCode uses a single opencode.json file in the project root for all configuration, including MCP servers.
Quick setup
Section titled “Quick setup”cce init # Auto-detects OpenCode if opencode.json existscce init --agent all # Explicitly includes OpenCodeFiles created
Section titled “Files created”opencode.json
Section titled “opencode.json”CCE adds its MCP server entry to the existing opencode.json (or creates one if it does not exist).
{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}No instruction file
Section titled “No instruction file”OpenCode does not use a separate instruction file. The MCP server registration is sufficient for OpenCode to discover and use CCE’s tools.
+Auto-detection
Section titled “Auto-detection”CCE detects OpenCode when an opencode.json file exists in your project root. No explicit --agent flag is needed.
Multi-Agent Support
Code Context Engine works with any AI coding agent that supports MCP (Model Context Protocol). The cce init command auto-detects which agents are present in your environment and configures them automatically.
The --agent flag
Section titled “The --agent flag”cce init --agent auto # Default. Detects installed agents.cce init --agent claude # Configure only Claude Codecce init --agent cursor # Configure only Cursorcce init --agent copilot # Configure only VS Code / Copilotcce init --agent gemini # Configure only Gemini CLIcce init --agent codex # Configure only Codex CLIcce init --agent all # Configure all supported agentsWhen no --agent flag is provided, cce init defaults to auto, which scans for known config files and editors.
Supported Editors and Agents
Section titled “Supported Editors and Agents”| Agent | MCP Config Path | Instruction File |
|---|---|---|
| Claude Code | .mcp.json | CLAUDE.md |
| Cursor | .cursor/mcp.json | .cursorrules |
| VS Code / Copilot | .vscode/mcp.json | .github/copilot-instructions.md |
| Gemini CLI | .gemini/settings.json | GEMINI.md |
| Codex CLI | ~/.codex/config.toml (global) | AGENTS.md |
| OpenCode | opencode.json | (none) |
| Tabnine | .tabnine/agent/settings.json | TABNINE.md |
How it works
Section titled “How it works”Each agent integration does two things:
+-
+
- Registers the MCP server so the agent can call
context_searchand other CCE tools.
+ - Writes an instruction file telling the agent to prefer CCE’s search over raw file reads. +
The instruction file content is managed by CCE and wrapped in markers (CCE:BEGIN / CCE:END) so it can be updated on upgrade without touching your own content.
Re-running for additional agents
Section titled “Re-running for additional agents”You can run cce init --agent <name> multiple times. Each run is additive and will not remove previously configured agents.
cce init --agent claudecce init --agent copilot # Adds Copilot config alongside ClaudeOr configure everything at once:
+cce init --agent allTabnine
Tabnine uses a project-local settings file and an instruction file for MCP integration.
+Quick setup
Section titled “Quick setup”cce init # Auto-detects Tabnine if .tabnine/ existscce init --agent all # Explicitly includes TabnineFiles created
Section titled “Files created”.tabnine/agent/settings.json
Section titled “.tabnine/agent/settings.json”Registers the CCE MCP server for Tabnine’s agent.
+{ "mcpServers": { "context-engine": { "command": "cce", "args": ["serve"] } }}TABNINE.md
Section titled “TABNINE.md”Contains instructions for Tabnine to prefer context_search for code retrieval. The CCE block is wrapped in markers so your own content is preserved.
Auto-detection
Section titled “Auto-detection”CCE detects Tabnine when a .tabnine/ directory exists in your project root. No explicit --agent flag is needed.
CLI Reference
cce init
Section titled “cce init”One-time setup for a project. Checks dependencies, indexes all code, installs git hooks, and connects AI coding agents via MCP.
+cce initcce init --agent claudecce init --agent copilotcce init --agent codexcce init --agent allWhat it does:
+-
+
- Downloads the embedding model (first run only, ~60 MB). +
- Checks Ollama status and reports compression mode. +
- Builds vector, FTS, and graph indexes. +
- Installs
post-commit,post-checkout, andpost-mergegit hooks.
+ - Writes MCP config for selected agents. +
- Creates or updates agent instruction files. +
- Adds per-machine files to
.gitignore.
+
cce index
Section titled “cce index”Re-index files that have changed since the last run.
+cce index # Incremental (changed files only)cce index --full # Force full re-index of every filecce index --path src/ # Index a specific file or directorycce index -v # Verbose outputThe git hooks installed by cce init call cce index automatically after every commit.
cce status
Section titled “cce status”Show index health and token savings summary.
+cce status # Full statuscce status --oneline # Single line (used by SessionStart hook)cce status --json # Machine-readable outputcce status -v # Lists all indexed projectscce savings
Section titled “cce savings”Token savings report with cost estimates.
+cce savings # Current projectcce savings --all # All indexed projectscce savings --json # Machine-readable outputcce search
Section titled “cce search”Run a test query against the index and display results.
+cce search 'how does authentication work'cce search 'payment processing' --top-k 10Also updates savings stats, useful for populating the dashboard before opening an agent session.
+cce dashboard
Section titled “cce dashboard”Open the web dashboard in your browser.
+cce dashboardcce dashboard --port 8080cce dashboard --no-browserThe dashboard provides views for: overview, files, sessions, and savings.
+cce services
Section titled “cce services”Manage Ollama and the dashboard as background processes.
+cce services # Show statuscce services start # Start Ollama + dashboardcce services start ollama # Start only Ollamacce services start dashboard # Start dashboardcce services start dashboard --port 9000cce services stop # Stop everything CCE startedcce services stop dashboard # Stop only dashboardcce start / cce stop
Section titled “cce start / cce stop”Shortcuts for cce services start and cce services stop.
cce start # Start all servicescce stop # Stop all servicescce start ollama # Start only Ollamacce stop dashboard # Stop only dashboardcce commands
Section titled “cce commands”Manage per-project rules, preferences, and shell hooks.
+cce commands list # Show all rules and hookscce commands add-rule 'Use UUID for PKs' # Add a rulecce commands remove-rule 'Use UUID for PKs'cce commands set-pref database PostgreSQL # Set a preferencecce commands remove-pref databasecce commands add before_push 'npm test' # Add hook commandcce commands remove before_push 'npm test'cce commands add-custom deploy 'kubectl apply -f k8s/'cce clear
Section titled “cce clear”Clear all index data for the current project.
+cce clear # Asks for confirmationcce clear --yes # Skip confirmationAfter clearing, run cce index --full to rebuild.
cce prune
Section titled “cce prune”Remove index data for projects whose directories no longer exist on disk.
+cce prune # Remove stale project datacce prune --dry-run # Preview without deletingcce upgrade
Section titled “cce upgrade”Upgrade CCE to the latest version. Detects your install method (uv, pipx, or pip) and runs the correct upgrade command. Refreshes project config afterwards.
+cce upgrade # Upgrade and refresh configcce upgrade --check # Show install method without upgradingcce uninstall
Section titled “cce uninstall”Remove CCE from the current project. Reverses everything cce init did.
cce uninstallRemoves: git hooks, MCP config entry, instruction file block, and .cce/ directory. Index data in ~/.cce is preserved (use cce clear to remove it).
cce serve
Section titled “cce serve”Start the MCP server. Called automatically by agents via .mcp.json. You do not need to run this manually.
cce servecce serve --project-dir /path/to/projectcce list
Section titled “cce list”Show every available command grouped by category.
+cce listOther flags
Section titled “Other flags”cce --version # Show versioncce --help # Show helpConfiguration
CCE works with zero configuration out of the box. This page covers all available options for when you need to tune behavior.
+Config file locations
Section titled “Config file locations”-
+
- Global:
~/.cce/config.yaml(created automatically on first use)
+ - Per-project:
.context-engine.yamlin your project root (overrides global for that project)
+
Full config reference
Section titled “Full config reference”compression: level: standard # How much to compress code chunks before sending to the agent # Options: minimal | standard | full output: standard # How much to compress agent responses # Options: off | lite | standard | max model: phi3:mini # Ollama model for LLM-based summarization # Auto-detected if Ollama is running. Ignored if Ollama is off.
+indexer: watch: true # Keep index in sync via git hooks ignore: # Directories and patterns to skip during indexing - .git - node_modules - __pycache__ - .venv - dist - build
+retrieval: top_k: 20 # Maximum chunks returned per query confidence_threshold: 0.5 # Minimum score to include a result (0.0 to 1.0)
+embedding: model: BAAI/bge-small-en-v1.5 # Embedding model (fastembed-compatible)
+pricing: model: opus # Model for cost estimates in `cce savings` # Options: opus | sonnet | haikuCompression levels
Section titled “Compression levels”Input compression (compression.level)
Section titled “Input compression (compression.level)”Controls how much CCE compresses code chunks before including them in the agent’s context.
+ + + + + + + + + + + + + + + + + + + + + +| Level | Behavior |
|---|---|
minimal | Truncation only. Keeps signature and docstring, drops body. |
standard | Truncation plus light summarization if Ollama is available. |
full | Full LLM summarization via Ollama (requires Ollama running). |
Output compression (compression.output)
Section titled “Output compression (compression.output)”Controls how verbose the agent’s responses are. Set via the set_output_compression MCP tool or config.
| Level | Style | Typical savings |
|---|---|---|
off | Full output | 0% |
lite | Removes filler and hedging | ~30% |
standard | Shorter phrasing, fragments where possible | ~65% |
max | Telegraphic, minimal prose | ~75% |
Code blocks, file paths, commands, and error messages are never compressed regardless of level.
+Change at runtime by telling your agent:
+Switch to max output compressionTurn off output compressionEmbedding model
Section titled “Embedding model”embedding: model: sentence-transformers/all-mpnet-base-v2Any model available in fastembed works. Changing the model requires a full re-index:
+cce clear --yes && cce index --fullThe default BAAI/bge-small-en-v1.5 is recommended for most use cases. It balances quality, speed, and size well.
Retrieval tuning
Section titled “Retrieval tuning”top_k controls how many chunks the retriever returns per query. Higher values surface more context but cost more tokens. Default: 20.
confidence_threshold sets the minimum score to include a result. Range 0.0 to 1.0. Lower values return more results; higher values return only strong matches. Default: 0.5.
At runtime, the agent can pass top_k and max_tokens directly to context_search:
context_search(query="payment processing", top_k=5, max_tokens=3000)Ignoring files
Section titled “Ignoring files”The indexer.ignore list supports:
-
+
- Directory names:
node_modules,dist
+ - File patterns:
"*.generated.ts","*.min.js"
+ - Relative paths:
"src/legacy/"
+
Files matching .gitignore are also skipped automatically.
Pricing model
Section titled “Pricing model”pricing: model: sonnet # opus (default) | sonnet | haikuThis determines which model’s pricing is used for cost estimates in cce savings. Prices are fetched from Anthropic’s docs and cached for 7 days.
Ollama URL
Section titled “Ollama URL”If Ollama is running on a non-default address, set it via environment variable:
+export OLLAMA_HOST=http://localhost:11434Resource profiles
Section titled “Resource profiles”CCE auto-detects available RAM and adjusts behavior:
+ + + + + + + + + + + + + + + + + + + + + + + + + +| RAM | Profile | Behavior |
|---|---|---|
| Less than 12 GB | light | Truncation only, small embedding batches |
| 12 to 32 GB | standard | Full pipeline, standard batch sizes |
| More than 32 GB | full | Larger Ollama models, larger batches |
You do not need to set this manually.
+Security
Section titled “Security”-
+
- All data stays local. No code is sent to external services (unless you use a cloud embedding model). +
- Index data is stored in
~/.cce/projects/.
+ - The MCP server only listens on stdio (not network) when launched by an agent. +
FAQ
Does CCE affect answer quality?
Section titled “Does CCE affect answer quality?”No. CCE returns the same code your agent would find by reading files, just compressed and targeted. In practice, answers are often better because the agent receives focused, relevant context instead of entire files full of unrelated code.
+How can I increase output savings?
Section titled “How can I increase output savings?”Set output compression to a higher level:
+compression: output: maxOr tell your agent at runtime: “Switch to max output compression.” The max level uses telegraphic phrasing and typically saves ~75% on response tokens. Code blocks and file paths are never affected.
Where do the savings come from?
Section titled “Where do the savings come from?”Three main sources:
+-
+
- Retrieval. Only relevant chunks are returned instead of the full codebase. This is the largest contributor (often 80%+ reduction). +
- Chunk compression. Retrieved chunks are truncated to signatures and docstrings, or summarized via Ollama if available. +
- Output compression. Agent responses are shortened by removing filler, hedging, and verbose phrasing. +
Is my code sent anywhere?
Section titled “Is my code sent anywhere?”No. All processing happens locally:
+-
+
- Embedding uses a local model downloaded to your machine. +
- Vector search runs in a local SQLite database. +
- The MCP server communicates over stdio (not network). +
- Ollama summarization (if enabled) also runs locally. +
No code, embeddings, or queries leave your machine unless you explicitly configure a remote embedding model.
+Does it work offline?
Section titled “Does it work offline?”Yes, fully. After the initial setup (which downloads the embedding model, ~60 MB), CCE operates entirely offline. Ollama summarization also runs locally if you have it installed.
+The only network call CCE makes is fetching model pricing for cost estimates in cce savings, and that result is cached for 7 days.
What languages are supported?
Section titled “What languages are supported?”CCE uses Tree-sitter for structural parsing. The following languages have full AST-aware chunking:
+-
+
- Python +
- JavaScript +
- TypeScript +
- PHP +
- Go +
- Rust +
- Java +
Other file types (YAML, Markdown, config files, etc.) are indexed using line-based chunking. They still appear in search results but without function-level granularity.
+Can I use CCE with multiple agents at once?
Section titled “Can I use CCE with multiple agents at once?”Yes. Run cce init --agent all to configure every supported agent. They all share the same index and MCP server, so there is no duplication or conflict.
How do I update CCE?
Section titled “How do I update CCE?”cce upgradeThis detects your install method (uv, pipx, or pip), upgrades the package, and refreshes your project config (hooks, MCP config, instruction files).
+How do I remove CCE from a project?
Section titled “How do I remove CCE from a project?”cce uninstallThis removes git hooks, MCP config entries, instruction file blocks, and the local .cce/ directory. Index data in ~/.cce is preserved. Run cce clear afterwards to remove that too.
The savings show 0 queries. What’s wrong?
Section titled “The savings show 0 queries. What’s wrong?”Savings are recorded when your agent calls context_search through the MCP server. If you have not used an agent session yet, run a test search to seed the stats:
cce search 'main entry point'Getting Started
System requirements
Section titled “System requirements”-
+
- Python 3.11+ (tested on 3.11, 3.12, 3.13) +
- A C compiler and
cmake(needed to build tree-sitter grammars)
+
| Platform | Setup |
|---|---|
| macOS | xcode-select --install |
| Ubuntu/Debian | sudo apt install build-essential cmake |
| Fedora/RHEL | sudo dnf install gcc gcc-c++ cmake |
| Windows | Visual Studio Build Tools (C++ workload) + CMake |
Install
Section titled “Install”uv tool install code-context-engineOr with pipx:
+pipx install code-context-engineOptional: Local embedding (no Ollama)
Section titled “Optional: Local embedding (no Ollama)”uv tool install "code-context-engine[local]" # includes fastembed + ONNX RuntimeInitialize your project
Section titled “Initialize your project”cd /path/to/your/projectcce initThis does everything:
+-
+
- Detects your embedding backend (Ollama or fastembed) +
- Builds vector, FTS, and graph indexes +
- Installs git hooks (auto-updates index on commit) +
- Writes MCP config for detected editors +
- Creates instruction files +
Target a specific agent
Section titled “Target a specific agent”cce init --agent claude # Claude Code onlycce init --agent codex # Codex CLI onlycce init --agent copilot # VS Code / Copilot onlycce init --agent all # Every supported editorVerify it works
Section titled “Verify it works”Restart your editor, then ask a question about your code. The agent will call context_search via MCP instead of reading files.
Check your savings:
+cce savings my-project · 5 queries
+ ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ 93% tokens saved
+ Input savings 42.1k tokens $0.63 Output savings 1.2k tokens $0.09 ────────────────────────────────────────── Total saved 43.3k tokens $0.72Embedding backends
Section titled “Embedding backends”CCE auto-detects the best available backend:
+-
+
- Ollama (preferred) — If running at localhost:11434, uses
nomic-embed-text. Zero extra dependencies.
+ - fastembed — Install with
[local]extra. UsesBAAI/bge-small-en-v1.5. Works offline, ~60 MB download.
+
Set CCE_EMBED_BACKEND=ollama or CCE_EMBED_BACKEND=fastembed to force a specific backend.
Next steps
Section titled “Next steps”-
+
- Multi-agent setup — Configure all your editors +
- Configuration — Tune compression, embedding, and more +
- CLI Reference — All available commands +
How It Works
CCE sits between your AI coding agent and your codebase. It replaces full-file reads with compressed, relevant chunks, reducing token usage while preserving answer quality.
+Indexing pipeline
Section titled “Indexing pipeline”When you run cce init or cce index, the following steps execute:
-
+
-
+
Tree-sitter parsing. Each source file is parsed into an AST using language-specific Tree-sitter grammars. This identifies functions, classes, methods, and other structural units.
+
+ -
+
Chunking. The AST is split into semantic chunks (one per function, class, or logical block). Each chunk retains its file path, line range, and relationships to other chunks.
+
+ -
+
Embedding. Each chunk is embedded using a local model (default:
+BAAI/bge-small-en-v1.5via fastembed). No data leaves your machine.
+ -
+
Storage. Embeddings, full-text content, and graph edges are written to a local SQLite database with sqlite-vec for vector search and FTS5 for keyword search.
+
+
Search pipeline
Section titled “Search pipeline”When an agent calls context_search, the following steps execute:
-
+
-
+
Query embedding. The natural language query is embedded using the same model.
+
+ -
+
Hybrid retrieval. Two searches run in parallel:
+-
+
- Vector similarity search (semantic match via sqlite-vec). +
- Full-text keyword search (BM25 via FTS5). +
+ -
+
RRF merge. Results from both searches are combined using Reciprocal Rank Fusion, which produces a single ranked list without needing score normalization.
+
+ -
+
Graph expansion. Top results are expanded by following code relationships (calls, imports, inheritance) to pull in related chunks the query might not have matched directly.
+
+ -
+
Compression. The final chunk set is compressed before being returned to the agent.
+
+
Storage
Section titled “Storage”All index data lives in ~/.cce/projects/<project-name>/:
-
+
- Vector index: sqlite-vec extension for approximate nearest neighbor search. +
- Full-text index: FTS5 for keyword/BM25 retrieval. +
- Graph: Edges representing code relationships (function calls, imports, class inheritance). +
- Metadata: File hashes for incremental indexing, chunk boundaries, and statistics. +
Everything is SQLite. No external database required.
+Compression layers
Section titled “Compression layers”CCE applies multiple compression stages to minimize tokens while preserving usefulness:
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +| Layer | What it does |
|---|---|
| Retrieval | Only relevant chunks are returned (not the whole codebase). |
| Chunk compression | Function bodies are truncated to signature + docstring, or summarized via Ollama. |
| Output compression | Agent responses are made more concise (configurable level). |
| Grammar compression | Removes syntactic noise (extra whitespace, redundant type annotations) from returned code. |
| Turn summarization | Long conversation histories are summarized to reduce context window usage. |
| Progressive disclosure | Returns signatures first; the agent can request full bodies only when needed. |
Supported languages
Section titled “Supported languages”Tree-sitter grammars are included for:
+-
+
- Python +
- JavaScript +
- TypeScript +
- PHP +
- Go +
- Rust +
- Java +
Other file types are indexed using line-based chunking without AST awareness.
Introduction
Code Context Engine (CCE) is a local MCP server that indexes your codebase so AI coding agents search for relevant code instead of reading entire files.
+The problem
Section titled “The problem”Every time an AI agent needs to understand your code, it reads entire files. A 500-line file costs 500 lines of input tokens even when the agent only needs one function. Across a session, this adds up to thousands of wasted tokens and real dollars.
+The solution
Section titled “The solution”CCE parses your code into semantic chunks (functions, classes, modules) using Tree-sitter, stores them with vector embeddings, and serves only the relevant pieces when the agent asks a question.
+Result: 94% input token savings, reproducibly benchmarked.
+What you get
Section titled “What you get”| Tool | Purpose |
|---|---|
context_search | Hybrid vector + keyword search with graph expansion |
get_chunk | Retrieve a specific chunk by ID |
record_decision | Store architectural decisions for cross-session recall |
record_code_area | Mark areas you’ve worked on |
session_recall | Recall decisions and code areas |
session_timeline | Browse tool call history |
session_event | Inspect a specific past event |
set_output_level | Control output compression (off/lite/standard/max) |
set_scope | Limit search to specific directories |
Supported agents
Section titled “Supported agents”| Editor | Config written | Instructions |
|---|---|---|
| Claude Code | .mcp.json | CLAUDE.md |
| VS Code / Copilot | .vscode/mcp.json | .github/copilot-instructions.md |
| Cursor | .cursor/mcp.json | .cursorrules |
| Gemini CLI | .gemini/settings.json | GEMINI.md |
| OpenAI Codex | ~/.codex/config.toml | AGENTS.md |
| OpenCode | opencode.json | |
| Tabnine | .tabnine/agent/settings.json | TABNINE.md |
How it works
Section titled “How it works”-
+
- Index — Tree-sitter parses code into semantic chunks. Stored locally with vector embeddings. +
- Search — Agent calls
context_searchvia MCP. Hybrid vector + BM25 merged with Reciprocal Rank Fusion. Graph expansion adds related imports.
+ - Compress — Chunks are compressed (truncation or LLM summary with Ollama). Output compression reduces reply tokens. +
- Track — Every query recorded.
cce savingsshows tokens and dollars saved.
+
${this.escapeHtml(e.details)}`:""} + `,this.appendChild(t),this.dispatchEvent(new CustomEvent("pagefind-error",{detail:e,bubbles:!0,composed:!0}))}escapeHtml(e){let t=document.createElement("div");return t.textContent=e,t.innerHTML}};var Re=class extends f{init(){this.setAttribute("hidden","")}register(s){s.registerUtility(this);let e=this.getAttribute("bundle-path");e&&(s.options.bundlePath=e);let t=this.getAttribute("base-url");t&&(s.pagefindOptions.baseUrl=t);let r=this.getAttribute("excerpt-length");r&&(s.pagefindOptions.excerptLength=parseInt(r,10));let i=this.getAttribute("lang");i&&s.setLanguage(i);let a=this.getAttribute("meta-cache-tag");a&&(s.pagefindOptions.metaCacheTag=a);let l=this.getAttribute("highlight-param");l&&(s.pagefindOptions.highlightParam=l),this.hasAttribute("exact-diacritics")&&(s.pagefindOptions.exactDiacritics=!0),this.hasAttribute("no-worker")&&(s.pagefindOptions.noWorker=!0),this.hasAttribute("faceted")&&(s.faceted=!0),this.hasAttribute("preload")&&s.triggerLoad()}};customElements.get("pagefind-config")||customElements.define("pagefind-config",Re);var Ea=(n=100)=>new Promise(s=>setTimeout(s,n)),ke=class extends f{constructor(){super();this.inputEl=null;this.clearEl=null;this.searchID=0;this.placeholder="";this.debounce=300;this.autofocus=!1}static get observedAttributes(){return["placeholder","debounce","autofocus"]}readAttributes(){this.hasAttribute("placeholder")&&(this.placeholder=this.getAttribute("placeholder")||""),this.hasAttribute("debounce")&&(this.debounce=parseInt(this.getAttribute("debounce")||"300",10)||300),this.hasAttribute("autofocus")&&(this.autofocus=this.hasAttribute("autofocus"))}init(){this.readAttributes(),this.render()}render(){this.innerHTML="";let e=this.instance.generateId("pfmod-input"),t=this.instance?.translate("search_label")||"Search this site",r=this.instance?.translate("clear_search")||"Clear",i=this.placeholder||this.instance?.translate("placeholder")||"Search";this.instance?.direction==="rtl"?this.setAttribute("dir","rtl"):this.removeAttribute("dir");let a=document.createElement("search");a.className="pf-input-wrapper",a.setAttribute("role","search"),a.setAttribute("aria-label",t);let l=document.createElement("label");l.setAttribute("for",e),l.setAttribute("data-pf-sr-hidden","true"),l.textContent=t,a.appendChild(l),this.inputEl=document.createElement("input"),this.inputEl.id=e,this.inputEl.className="pf-input",this.inputEl.setAttribute("type","search"),this.inputEl.setAttribute("autocomplete","off"),this.inputEl.setAttribute("autocapitalize","none"),this.inputEl.setAttribute("enterkeyhint","search"),this.inputEl.setAttribute("placeholder",i),this.autofocus&&this.inputEl.setAttribute("autofocus","autofocus");let o=this.instance.generateId("pf-input-hint"),c=this.instance?.translate("input_hint")||"Results will appear as you type",u=document.createElement("span");u.id=o,u.setAttribute("data-pf-sr-hidden","true"),u.textContent=c,this.inputEl.setAttribute("aria-describedby",o),a.appendChild(this.inputEl),a.appendChild(u),this.clearEl=document.createElement("button"),this.clearEl.className="pf-input-clear",this.clearEl.setAttribute("type","button"),this.clearEl.setAttribute("data-pf-suppressed","true"),this.clearEl.textContent=r,a.appendChild(this.clearEl),this.appendChild(a),this.setupEventHandlers()}setupEventHandlers(){!this.inputEl||!this.clearEl||(this.inputEl.addEventListener("input",async e=>{let t=e.target;if(this.instance&&typeof t?.value=="string"){this.updateState(t.value);let r=++this.searchID;if(await Ea(this.debounce),r!==this.searchID)return;this.instance?.triggerSearch(t.value)}}),this.inputEl.addEventListener("keydown",e=>{e.key==="Escape"&&(++this.searchID,this.inputEl&&(this.inputEl.value=""),this.instance?.triggerSearch(""),this.updateState("")),e.key==="ArrowDown"&&(e.preventDefault(),this.inputEl&&this.instance?.focusNextResults(this.inputEl))}),this.inputEl.addEventListener("focus",()=>{this.instance?.triggerLoad();let e=this.instance?.translate("keyboard_navigate")||"navigate",t=this.instance?.translate("keyboard_clear")||"clear";this.instance?.registerShortcut({label:"\u2193",description:e},this),this.instance?.registerShortcut({label:"esc",description:t},this)}),this.inputEl.addEventListener("blur",()=>{this.instance?.deregisterAllShortcuts(this)}),this.clearEl.addEventListener("click",()=>{this.inputEl&&(this.inputEl.value="",this.instance?.triggerSearch(""),this.updateState(""),this.inputEl.focus())}))}updateState(e){this.clearEl&&(e&&e?.length?this.clearEl.removeAttribute("data-pf-suppressed"):this.clearEl.setAttribute("data-pf-suppressed","true"))}register(e){e.registerInput(this,{keyboardNavigation:!0}),e.on("search",t=>{this.inputEl&&document.activeElement!==this.inputEl&&(this.inputEl.value=t,this.updateState(t))},this),e.on("error",t=>{let r=t;this.showError({message:r.message||"Search initialization failed",details:r.bundlePath?`Bundle path: ${r.bundlePath}`:void 0})},this),e.on("translations",()=>{let t=this.inputEl?.value||"";this.render(),this.inputEl&&t&&(this.inputEl.value=t,this.updateState(t))},this)}update(){this.render()}focus(){this.inputEl&&this.inputEl.focus()}};customElements.get("pagefind-input")||customElements.define("pagefind-input",ke);var Ae=class extends f{constructor(){super();this.containerEl=null;this.term="";this.defaultMessage=""}static get observedAttributes(){return["default-message"]}init(){this.hasAttribute("default-message")&&(this.defaultMessage=this.getAttribute("default-message")||""),this.render()}render(){this.innerHTML="",this.instance?.direction==="rtl"?this.setAttribute("dir","rtl"):this.removeAttribute("dir"),this.containerEl=document.createElement("div"),this.containerEl.className="pf-summary",this.containerEl.textContent=this.defaultMessage,this.appendChild(this.containerEl)}reconcileAria(){}register(e){e.registerSummary(this),e.on("search",t=>{this.term=t},this),e.on("results",t=>{if(!this.containerEl||!t)return;let i=t?.results?.length??0;if(!this.term){if(e.faceted){let o=i===0?"total_zero_results":i===1?"total_one_result":"total_many_results",c=e.translate(o,{COUNT:i});this.containerEl.textContent=c||`${i} result${i===1?"":"s"}`}else this.containerEl.textContent=this.defaultMessage;return}let a=i===0?"zero_results":i===1?"one_result":"many_results",l=e.translate(a,{SEARCH_TERM:this.term,COUNT:i});this.containerEl.textContent=l||`${i} result${i===1?"":"s"} for ${this.term}`},this),e.on("loading",()=>{if(!this.containerEl)return;let t=e.translate("searching",{SEARCH_TERM:this.term});this.containerEl.textContent=t||`Searching for ${this.term}...`},this),e.on("error",t=>{if(!this.containerEl)return;let r=t,i=e.translate("error_search")||"Search failed";this.containerEl.textContent=`Error: ${r.message||i}`},this),e.on("translations",()=>{this.render()},this)}update(){this.hasAttribute("default-message")&&(this.defaultMessage=this.getAttribute("default-message")||"",!this.term&&this.containerEl&&(this.containerEl.textContent=this.defaultMessage))}};customElements.get("pagefind-summary")||customElements.define("pagefind-summary",Ae);var ss=n=>{if(n instanceof Element)return[n];if(Array.isArray(n)&&n.every(s=>s instanceof Element))return n;if(typeof n=="string"||n instanceof String){let s=document.createElement("div");return s.innerHTML=n,[...s.childNodes]}return console.error(`[Pagefind Results]: Expected template to return HTML element or string, got ${typeof n}`),[]},ba=`
+ {{#if and(options.show_images, meta.image)}}
+ ) }})
+ {{/if}}
+
+ {{#if sub_results}}
+
+
+ + {{ meta.title }} +
+ {{#if excerpt}} +{{+ excerpt +}}
+ {{/if}} +-
+ {{#each sub_results as sub}}
+
-
+ {{ sub.title }}
+
{{+ sub.excerpt +}}
+
+ {{/each}}
+
{{/if}}
+
{{/if}}`,ns=R(Ra),ka=``,Aa=R(ka),jt=class{constructor(s){this.data=null;this.cachedOptions=null;this.loading=!1;this.retryDelay=0;this.observer=null;this.rawResult=s.rawResult,this.placeholderEl=s.placeholderEl,this.renderFn=s.renderFn,this.intersectionRoot=s.intersectionRoot,this.index=s.index,this.onLoad=s.onLoad,this.setupObserver()}setupObserver(){if(this.data!==null||this.observer!==null)return;let s={root:this.intersectionRoot,rootMargin:"50px",threshold:.01};this.observer=new IntersectionObserver((e,t)=>{this.data===null&&e?.[0]?.isIntersecting&&(this.load(),t.disconnect(),this.observer=null)},s),this.observer.observe(this.placeholderEl)}async load(){if(!(this.data!==null||this.loading)){this.loading=!0;try{this.data=await this.rawResult.data();let s=this.renderFn(this.data),e=Ke(s);if(e.length>0&&this.placeholderEl.parentNode){let t=e.find(r=>r instanceof Element);this.placeholderEl.replaceWith(...e),t instanceof Element&&(this.placeholderEl=t,as(t,this.index),this.cacheOptions())}}catch{await new Promise(s=>setTimeout(s,this.retryDelay||100)),this.retryDelay=Math.min((this.retryDelay||100)*2,1e4),this.loading=!1}this.onLoad?.()}}cacheOptions(){if(!this.data||!this.placeholderEl){this.cachedOptions=null;return}this.placeholderEl.getAttribute("role")==="group"?this.cachedOptions=Array.from(this.placeholderEl.querySelectorAll('[role="option"]')):this.placeholderEl.getAttribute("role")==="option"?this.cachedOptions=[this.placeholderEl]:this.cachedOptions=[]}cleanup(){this.observer&&(this.observer.disconnect(),this.observer=null),this.cachedOptions=null}},Fe=class extends f{constructor(){super();this.containerEl=null;this.inputEl=null;this.dropdownEl=null;this.resultsEl=null;this.statusEl=null;this.footerEl=null;this.isOpen=!1;this.isLoading=!1;this.results=[];this.activeIndex=-1;this.activeOptionOffset=0;this.searchID=0;this.searchTerm="";this.pendingNavigation=0;this.loadingAnnouncementTimeout=null;this.selectedEl=null;this._userPlaceholder=null;this.debounce=150;this.autofocus=!1;this.showSubResults=!1;this.maxResults=0;this.showKeyboardHints=!0;this.shortcut="mod+k";this.hideShortcut=!1;this.resultTemplate=null;this.compiledResultTemplate=null;this.compiledPlaceholderTemplate=null;this._documentClickHandler=null;this._shortcutKeyHandler=null;this._keyBinding=null;this._shortcutEl=null}static get observedAttributes(){return["placeholder","debounce","autofocus","show-sub-results","max-results","show-keyboard-hints","shortcut","hide-shortcut"]}get placeholder(){return this._userPlaceholder||this.instance?.translate("placeholder")||"Search..."}readAttributes(){this.hasAttribute("placeholder")&&(this._userPlaceholder=this.getAttribute("placeholder")),this.hasAttribute("debounce")&&(this.debounce=parseInt(this.getAttribute("debounce")||"150",10)||150),this.hasAttribute("autofocus")&&(this.autofocus=this.hasAttribute("autofocus")),this.hasAttribute("show-sub-results")&&(this.showSubResults=this.getAttribute("show-sub-results")!=="false"),this.hasAttribute("max-results")&&(this.maxResults=parseInt(this.getAttribute("max-results")||"0",10)),this.hasAttribute("show-keyboard-hints")&&(this.showKeyboardHints=this.getAttribute("show-keyboard-hints")!=="false"),this.hasAttribute("shortcut")&&(this.shortcut=this.getAttribute("shortcut")||"mod+k"),this.hasAttribute("hide-shortcut")&&(this.hideShortcut=this.getAttribute("hide-shortcut")!=="false"),this._keyBinding=je(this.shortcut)}init(){this.readAttributes(),this.checkForTemplates(),this.render(),this.setupOutsideClickHandler(),this.setupShortcutHandler()}checkForTemplates(){let e=this.querySelector('script[type="text/pagefind-template"]:not([data-template]), script[type="text/pagefind-template"][data-template="result"]');e&&(this.compiledResultTemplate=R((e.textContent||"").trim()));let t=this.querySelector('script[type="text/pagefind-template"][data-template="placeholder"]');t&&(this.compiledPlaceholderTemplate=R((t.textContent||"").trim()))}getPlaceholder(){return this.compiledPlaceholderTemplate?this.compiledPlaceholderTemplate({}):Aa({})}render(){let e=[];this.querySelectorAll('script[type="text/pagefind-template"]').forEach(l=>{e.push(l)}),this.innerHTML="",e.forEach(l=>this.appendChild(l));let t=this.instance.generateId("pf-sb-input"),r=this.instance.generateId("pf-sb-results");this.containerEl=document.createElement("div"),this.containerEl.className="pf-searchbox",this.appendChild(this.containerEl);let i=document.createElement("div");if(i.className="pf-searchbox-input-wrapper",this.containerEl.appendChild(i),this.inputEl=document.createElement("input"),this.inputEl.id=t,this.inputEl.className="pf-searchbox-input",this.inputEl.type="text",this.inputEl.setAttribute("role","combobox"),this.inputEl.setAttribute("aria-autocomplete","list"),this.inputEl.setAttribute("aria-controls",r),this.inputEl.setAttribute("aria-expanded","false"),this.inputEl.setAttribute("autocomplete","off"),this.inputEl.setAttribute("autocapitalize","none"),this.inputEl.placeholder=this.placeholder,this.autofocus&&this.inputEl.setAttribute("autofocus","autofocus"),i.appendChild(this.inputEl),!this.hideShortcut&&this._keyBinding){this._shortcutEl=document.createElement("span"),this._shortcutEl.className="pf-trigger-shortcut",this._shortcutEl.setAttribute("aria-hidden","true");let l=xe(this._keyBinding);for(let o of l.keys){let c=document.createElement("span");c.className="pf-trigger-key",c.textContent=o,this._shortcutEl.appendChild(c)}i.appendChild(this._shortcutEl),this.inputEl.setAttribute("aria-keyshortcuts",l.aria)}this.dropdownEl=document.createElement("div"),this.dropdownEl.className="pf-searchbox-dropdown",this.containerEl.appendChild(this.dropdownEl);let a=this.instance?.translate("results_label")||"Search results";this.instance?.direction==="rtl"?this.setAttribute("dir","rtl"):this.removeAttribute("dir"),this.resultsEl=document.createElement("div"),this.resultsEl.id=r,this.resultsEl.className="pf-searchbox-results",this.resultsEl.setAttribute("role","listbox"),this.resultsEl.setAttribute("aria-label",a),this.dropdownEl.appendChild(this.resultsEl),this.statusEl=document.createElement("div"),this.statusEl.className="pf-searchbox-status",this.statusEl.hidden=!0,this.dropdownEl.appendChild(this.statusEl),this.showKeyboardHints&&(this.footerEl=document.createElement("div"),this.footerEl.className="pf-searchbox-footer",this.footerEl.setAttribute("aria-hidden","true"),this.dropdownEl.appendChild(this.footerEl),this.renderFooterHints()),this.setupEventHandlers()}renderFooterHints(){if(!this.footerEl)return;this.footerEl.innerHTML="";let e=this.instance?.translate("keyboard_navigate")||"navigate",t=this.instance?.translate("keyboard_select")||"select",r=this.instance?.translate("keyboard_close")||"close",i=document.createElement("div");i.className="pf-searchbox-footer-hint";let a=document.createElement("span");a.className="pf-searchbox-footer-key",a.textContent="\u2191",i.appendChild(a);let l=document.createElement("span");l.className="pf-searchbox-footer-key",l.textContent="\u2193",i.appendChild(l),i.appendChild(document.createTextNode(` ${e}`)),this.footerEl.appendChild(i);let o=document.createElement("div");o.className="pf-searchbox-footer-hint";let c=document.createElement("span");c.className="pf-searchbox-footer-key",c.textContent="\u21B5",o.appendChild(c),o.appendChild(document.createTextNode(` ${t}`)),this.footerEl.appendChild(o);let u=document.createElement("div");u.className="pf-searchbox-footer-hint";let d=document.createElement("span");d.className="pf-searchbox-footer-key",d.textContent="esc",u.appendChild(d),u.appendChild(document.createTextNode(` ${r}`)),this.footerEl.appendChild(u)}setupEventHandlers(){!this.inputEl||!this.resultsEl||(this.inputEl.addEventListener("input",async e=>{let t=e.target.value;if(this.searchTerm=t,!t||!t.trim()){this.closeDropdown(),this.results=[],this.instance?.triggerSearch("");return}this.openDropdown(),this.showLoadingState();let r=++this.searchID;await va(this.debounce),r===this.searchID&&this.instance?.triggerSearch(t)}),this.inputEl.addEventListener("keydown",e=>{switch(e.key){case"ArrowDown":e.preventDefault(),!this.isOpen&&this.inputEl?.value.trim()&&this.openDropdown(),this.isOpen&&this.results.length>0&&this.moveSelection(1);break;case"ArrowUp":e.preventDefault(),this.isOpen&&this.results.length>0&&this.moveSelection(-1);break;case"Enter":this.isOpen&&this.activeIndex>=0?(e.preventDefault(),this.activateCurrentSelection(e)):!this.isOpen&&this.inputEl?.value.trim()&&(e.preventDefault(),this.openDropdown(),this.results.length>0?(this.rerenderLoadedResults(),this.activeIndex=0,this.activeOptionOffset=0,this.updateSelectionUI()):this.instance?.triggerSearch(this.inputEl.value));break;case"Escape":this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.isOpen&&(e.preventDefault(),this.closeDropdown());break;case"Tab":this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.isOpen&&this.closeDropdown();break}}),this.inputEl.addEventListener("focus",()=>{this.instance?.triggerLoad()}),this.resultsEl.addEventListener("click",e=>{e.target.closest("a")&&this.closeDropdown()}),this.resultsEl.addEventListener("mousemove",e=>{let t=e.target.closest("a");if(t){let r=this.getResultAndOffsetFromElement(t);r&&(r.resultIndex!==this.activeIndex||r.optionOffset!==this.activeOptionOffset)&&(this.activeIndex=r.resultIndex,this.activeOptionOffset=r.optionOffset,this.updateSelectionUI(!1))}}))}setupOutsideClickHandler(){this._documentClickHandler=e=>{this.isOpen&&!this.contains(e.target)&&this.closeDropdown()},document.addEventListener("click",this._documentClickHandler)}setupShortcutHandler(){this._keyBinding&&(this._shortcutKeyHandler=e=>{if(!this._keyBinding||!Be(this._keyBinding,e))return;let t=document.activeElement;t&&(t.tagName==="INPUT"||t.tagName==="TEXTAREA"||t.isContentEditable)||(e.preventDefault(),this.inputEl?.focus())},document.addEventListener("keydown",this._shortcutKeyHandler))}openDropdown(){this.isOpen||!this.containerEl||!this.inputEl||(this.isOpen=!0,this.containerEl.classList.add("open"),this.inputEl.setAttribute("aria-expanded","true"))}closeDropdown(){!this.isOpen||!this.containerEl||!this.inputEl||(this.isOpen=!1,this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.containerEl.classList.remove("open"),this.inputEl.setAttribute("aria-expanded","false"),this.inputEl.removeAttribute("aria-activedescendant"),this.activeIndex=-1,this.activeOptionOffset=0,this.selectedEl=null)}showLoadingState(){if(!this.resultsEl||!this.statusEl)return;this.isLoading=!0,this.resultsEl.innerHTML="",this.selectedEl=null,this.resultsEl.setAttribute("aria-busy","true");let e=this.instance?.translate("searching",{SEARCH_TERM:this.searchTerm})||"Searching...";this.statusEl.textContent=e,this.statusEl.className="pf-searchbox-status pf-searchbox-loading",this.statusEl.hidden=!1}showEmptyState(){if(!this.resultsEl||!this.statusEl)return;this.resultsEl.innerHTML="",this.selectedEl=null,this.resultsEl.removeAttribute("aria-busy");let e=this.instance?.translate("zero_results",{SEARCH_TERM:this.searchTerm})||`No results for "${this.searchTerm}"`;this.statusEl.textContent=e,this.statusEl.className="pf-searchbox-status pf-searchbox-empty",this.statusEl.hidden=!1,this.instance?.announce("zero_results",{SEARCH_TERM:this.searchTerm},"assertive")}getOptionsForResult(e){return e.cachedOptions!==null?e.cachedOptions:!e.data||!e.placeholderEl?[]:e.placeholderEl.getAttribute("role")==="group"?Array.from(e.placeholderEl.querySelectorAll('[role="option"]')):e.placeholderEl.getAttribute("role")==="option"?[e.placeholderEl]:[]}moveSelection(e){let t=this.results.length;if(t===0)return;if(e<0){if(this.activeIndex===-1)return;if(this.activeOptionOffset>0){this.activeOptionOffset--,this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.updateSelectionUI(!0);return}let o=this.activeIndex-1;if(o<0){this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.activeIndex=-1,this.activeOptionOffset=0,this.updateSelectionUI(!0);return}let c=this.results[o];if(!c||!c.data)return;let u=this.getOptionsForResult(c);this.activeIndex=o,this.activeOptionOffset=Math.max(0,u.length-1),this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.updateSelectionUI(!0),this.preloadAhead(o,e);return}if(this.activeIndex===-1){if(this.results[0]&&!this.results[0].data){this.pendingNavigation+=e,this.results[0].load(),this.scheduleLoadingAnnouncement(),this.preloadAhead(0,e);return}this.activeIndex=0,this.activeOptionOffset=0,this.pendingNavigation=0,this.clearLoadingAnnouncement(),this.updateSelectionUI(!0),this.preloadAhead(0,e);return}let r=this.results[this.activeIndex];if(!r?.data){r&&(this.pendingNavigation+=e,r.load(),this.scheduleLoadingAnnouncement(),this.preloadAhead(this.activeIndex,e));return}let i=this.getOptionsForResult(r);if(this.activeOptionOffset{{ meta.title | default("Untitled") }}
+ {{#if excerpt}} +{{+ excerpt +}}
+ {{/if}} +{{#if and(options.show_sub_results, sub_results)}} +{{#each sub_results as sub}} + +{{ sub.title | default("Section") }}
+ {{#if sub.excerpt}} +{{+ sub.excerpt +}}
+ {{/if}} + +{{/each}} +
+
+${i(30)}
+${i(40)}
+Savings Tracking
CCE tracks every query made through the MCP server and records how many tokens were served versus how many would have been needed without CCE. This data powers the cce savings command and the dashboard.
Using cce savings
Section titled “Using cce savings”cce savingsExample output:
+ my-project · 42 queries
+ ⛁ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ ⛶ 93% tokens saved
+ Without CCE 48.0k tokens $0.24 With CCE 3.4k tokens $0.02 ────────────────────────────────────────── Saved 44.6k tokens $0.22 ~81 tokens / query ~<$0.01 / query
+ How: retrieval 93% + compression 90% Cost estimate based on Opus input pricing ($5/1M tokens)Understanding the input/output split
Section titled “Understanding the input/output split”Savings come from two independent stages:
+-
+
-
+
Retrieval savings (input). Instead of sending the entire codebase, CCE returns only the chunks relevant to the query. This is measured as:
+1 - (served_tokens / full_codebase_tokens).
+ -
+
Compression savings (input). The retrieved chunks are further compressed (truncation, summarization) before being sent to the agent. This is measured as:
+1 - (compressed_tokens / raw_chunk_tokens).
+
The combined effect is multiplicative. If retrieval cuts 90% and compression cuts another 50%, the total savings are 95%.
+Per-bucket breakdown
Section titled “Per-bucket breakdown”The How: line in the output shows the contribution of each stage:
How: retrieval 93% + compression 90%-
+
- retrieval represents the savings from selecting only relevant chunks. +
- compression represents the savings from compressing those chunks. +
Configuring the pricing model
Section titled “Configuring the pricing model”Cost estimates use model-specific input pricing. Configure which model to estimate for:
+# ~/.cce/config.yaml or .context-engine.yamlpricing: model: opus # opus (default) | sonnet | haikuPrices are fetched from Anthropic’s documentation and cached for 7 days.
+Using cce dashboard
Section titled “Using cce dashboard”cce dashboardThe dashboard opens in your browser and provides a visual view of:
+-
+
- Total tokens saved over time (line chart). +
- Per-query breakdown. +
- Compression level controls (change input/output compression live). +
- File staleness detection. +
Cross-project savings
Section titled “Cross-project savings”cce savings --allShows a combined report across every project you have indexed, useful for understanding total cost reduction.
+JSON output
Section titled “JSON output”cce savings --jsonReturns machine-readable data for integration with other tools:
+{ "project": "my-project", "queries": 42, "served_tokens": 14200, "raw_tokens": 26000, "full_file_tokens": 48000, "tokens_saved": 33800, "savings_pct": 70, "retrieval_savings_pct": 46, "compression_savings_pct": 45}Populating savings before a session
Section titled “Populating savings before a session”If you have zero queries recorded (fresh install), run a test search to seed the stats:
+cce search 'how does the main module work'This updates the savings tracker so cce status and the dashboard show non-zero values.