Llamatik Code

The AI pair programmer built for Kotlin Multiplatform — works with every language and project in your JetBrains IDE.

Kotlin Multiplatform · Local by default · Distributed by choice

---

✨ What is Llamatik Code?

Llamatik Code is an open-source AI coding assistant built specifically for Kotlin Multiplatform development, running entirely on local inference. It wraps the Llamatik on-device AI SDK and gives you a KMP-aware agent that understands your shared/platform source sets, diagnoses native compilation errors, plans cross-platform refactors, and applies safe diffs — all inside your JetBrains IDE with no cloud dependency.

It works equally well for any JVM, Android, iOS, or general-purpose project — the KMP tooling is additive, not restrictive.

• 🎯 KMP-native — dedicated tools for diagnosing and fixing Kotlin Multiplatform native compilation issues (kmp_diagnose, kmp_fix_plan)
• 🤖 Agent-driven — a tool-calling loop with 70+ built-in tools for git, PSI, file I/O, search, build, and more
• 🔒 Privacy-first — inference runs locally by default via GGUF models and llama.cpp; you control what connects to the outside world
• 🧩 MCP-compatible — connect any Model Context Protocol server and let the agent call its tools
• 🩺 Code Health Monitor — background analysis for regressions, complexity, and refactor opportunities
• 🌐 External Knowledge — web search, docs research, and error analysis from inside the IDE with smart routing to kotlinlang.org, developer.android.com, and developer.apple.com
• 🚀 Autonomous Local Development — mission-driven coding: plan → patch → verify → audit, entirely on-device

---

🚀 Features

🤖 Autonomous Local Development (1.8.0+)

• Offline Enterprise Pair Programmer — describe a task; the agent classifies it, detects risks, plans, patches, verifies, and produces an audit report
• 11 auto-detected mission types — BUG_FIX, REFACTOR, FEATURE_IMPLEMENTATION, TEST_GENERATION, CODE_REVIEW, DOCUMENTATION, BUILD_FAILURE_REPAIR, ARCHITECTURE_EXPLANATION, DEPENDENCY_UPDATE, SECURITY_AUDIT, PERFORMANCE_ANALYSIS
• Safety levels — APPROVAL_REQUIRED (default), PLAN_ONLY, AUTO_APPLY, FULL_AUTO
• Risk detection — LOW / MEDIUM / HIGH / CRITICAL risks surfaced before any file is written; CRITICAL requires explicit acknowledgement
• Plan-before-write enforcement — no file is touched until a mission plan exists (configurable)
• Diff-before-write enforcement — unified diff shown before every patch (configurable)
• Audit trail — per-mission metadata log; accessible via the enterprise_mission_audit tool
• Inline completions — local fill-in-the-middle suggestions as you type (opt-in, debounced)
• Context menu — right-click any file or selection → Llamatik Code → Autonomous Local Development

💬 Agent & Chat

• ✅ Docked Llamatik Code tool window with Ask and Agent modes
• ✅ Multi-tab persistent chat sessions with per-project isolation
• ✅ Tool-driven agent loop with up to 5 execution rounds
• ✅ Safe diff preview before every file write
• ✅ Build verification after apply, with automatic error injection
• ✅ Agent trace panel — every tool call, duration, and result
• ✅ @filename attachment — pre-seeds files into context, no wasted read rounds
• ✅ Aggressive mode (Pro) — decisive agent runs without per-step confirmation
• ✅ Multi-task agent (Pro) — parallel tasks with a live task queue panel

🛠️ Built-in Tools (70+)

Core file tools

Tool	Description
read_file	Read the content of any project file
write_file	Write or overwrite a file (always previewed)
create_file	Create a new file
replace_in_file	Targeted in-file replacement
list_files	List files in a directory
search_code	Full-text search across the project
search_filenames	File search by name pattern
run_terminal	Execute a shell command
get_diagnostics	Fetch current IDE inspection / compiler errors
explain_build_failure	Interpret the last build failure

Git tools (1.3.0+)

Tool	Description
git_diff	Diff of current changes (staged + unstaged)
git_status	Working-tree status
git_log	Recent commit history
git_blame	Line-level blame for a file
git_changed_files	Files changed in the current diff

Project intelligence (1.3.0+)

Tool	Description
index_search_files	Fast file search by name pattern
index_search_symbols	Symbol search across the project index
index_list_recent	Recently opened or modified files
psi_find_usages	Find all usages of a symbol via PSI
psi_get_type	Resolve the type of an expression
psi_get_declarations	List declarations in a file
smart_search	Fuzzy search ranked by recency and relevance

MCP tools (1.4.0+)

Tool	Description
mcp_list_servers	List configured MCP servers
mcp_list_tools	Discover tools exposed by a server
mcp_call_tool	Call any MCP tool with approval gating
mcp_list_resources	List server resources
mcp_read_resource	Read a resource from an MCP server
mcp_list_prompts	List server prompt templates

Autonomous Local Development tools (1.8.0+)

Tool	Description
enterprise_mission_create	Start a structured coding mission
enterprise_mission_context	Gather project context for a mission
enterprise_mission_audit	Read the audit log for past missions
enterprise_mission_report	Produce the final mission report

🔌 MCP Server Support (1.4.0+)

• IDE-native MCP host — add any MCP server under Settings → Llamatik Code → MCP Servers
• Full server lifecycle: add, edit, remove, enable/disable, test connection, view logs
• Discovered MCP tools bridged as mcp_<server>_<tool> first-class agent tools
• Approval dialog with LOW / MEDIUM / HIGH risk level on every call
• HIGH-risk tools are never auto-approved
• Free: 1 active server, manual approval. Pro: unlimited servers + per-tool auto-approval rules
• Secrets redacted from all log and trace output

🌐 External Knowledge (1.5.0+)

• Web search via DuckDuckGo, documentation pages, GitHub Issues, and more
• Smart source routing to kotlinlang.org, developer.android.com, developer.apple.com, spring.io
• Error research: detects build tool and framework, produces structured fix plans
• Private docs indexing: add up to 10 (Pro) or 3 (Free) custom documentation sources
• Citations with every external research result (source type, confidence, URL)
• Free: 15 searches/day, 5 pages/task. Pro: 50 searches/day, 10 pages/task

🩺 Code Health Monitor (1.5.0+)

• Background analysis for large files, TODO/FIXME accumulation, empty catch blocks, deprecated APIs
• Regression risk analyser for git diffs: removed public APIs, deleted tests, dependency changes
• Watched areas: pin files, directories, or modules for priority analysis
• Before-commit assistant: summarises staged changes and flags regression risks
• Refactor opportunity detection: long functions, repeated catch patterns
• 14 dedicated agent tools for all Code Health operations
• Free: LIGHT mode, manual trigger only. Pro: BALANCED + AGGRESSIVE modes with background scheduling

✏️ Context Actions

Right-click any file or selection for one-shot operations:

• Ask About Selection · Explain This File · Refactor Selection · Fix Selected Code
• Generate Tests · Add Documentation · Edit with Prompt · Explain Failing Build
• Explain Current Git Changes · Autonomous Local Development submenu (1.8.0+)

---

📱 Try it now

Install from the JetBrains Plugin Marketplace — no setup beyond pointing the plugin at a GGUF model file.

---

🔧 Requirements

• IDE: IntelliJ IDEA 2026.1 or any JetBrains IDE on platform 252+ (IntelliJ IDEA, Android Studio, GoLand, CLion, PyCharm, WebStorm, and more)
• Model: Any instruction-tuned GGUF model compatible with llama.cpp (3B–34B+ depending on hardware)
• OS: macOS, Linux, or Windows (64-bit, JDK 21+)

---

⚡ Quick Start

1. Install the plugin from the JetBrains Marketplace (search Llamatik Code).
2. Restart your IDE.
3. Open Settings → Llamatik Code → Inference and set the Model path to a GGUF file on disk.
4. Open the Llamatik Code tool window (right panel) and start chatting.

Choosing a model

Any GGUF model compatible with llama.cpp works. Recommended starting points:

Use case	Model size	Quantisation
Fast / lightweight	3B – 7B	Q4_K_M or Q5_K_M
Balanced	13B – 14B	Q4_K_M
High quality	34B+	Q4_K_M (if hardware allows)

For agent and tool-use tasks, use instruction-tuned variants (names ending in -Instruct or -Chat).

Models can be downloaded from Hugging Face — search for GGUF quantisations of Qwen3, Gemma3, Mistral, LLaMA, or Phi.

---

🧑‍💻 Usage

Ask mode vs Agent mode

The tool window has two primary modes:

• Ask mode — Conversational Q&A. The model reads context but does not write files.
• Agent mode — Task execution. The agent calls tools, edits files, runs builds, and iterates.

Running a mission (Autonomous Local Development)

Type:  "Fix the build failure in AuthModule"
       "Refactor UserRepository without changing behavior"
       "Generate tests for PaymentService"
       "Review current git changes before committing"

The agent automatically detects the mission type, surfaces risks, and presents a plan for your approval before writing any files.

Alternatively, right-click any file → Llamatik Code → Autonomous Local Development → choose an action.

Attaching files

Use @filename in your prompt to attach a file. The agent pre-seeds it into context — no wasted read rounds.

Codebase rules

Use the rules_set tool (or ask the agent) to write persistent natural-language rules:

"Always use Result<T> for error returns. Never throw exceptions in repository classes."

Rules are injected into every agent session automatically.

---

🔍 Why Llamatik Code?

• ✅ Built for Kotlin Multiplatform — dedicated KMP tools, smart routing to KMP docs, awareness of shared/platform source sets
• ✅ Local by default — no cloud, no telemetry, no data exfiltration
• ✅ Built on Llamatik and llama.cpp
• ✅ 70+ built-in tools including git, PSI, build, and MCP
• ✅ Open-source (Apache 2.0)
• ✅ MCP-compatible — extend the agent with any tool server
• ✅ Structured mission workflow with plan approval and audit trail
• ✅ Works with any GGUF model — no vendor lock-in
• ✅ Works for any language or project in IntelliJ-based IDEs

---

📦 Benchmark harness

The project includes a headless benchmark suite that runs the real agent pipeline against self-contained fixture repositories and verifies the result by executing the fixture's own test suite.

Run all tasks (requires a model):

./gradlew runLlamatikBenchmark -PbenchmarkModel=<path-to-model>

Run with the scripted backend (no model needed — for CI smoke-testing):

./gradlew runLlamatikBenchmark -PbenchmarkModel=scripted

Run a single task:

./gradlew runLlamatikBenchmark -PbenchmarkModel=scripted -PbenchmarkTask=kotlin-simple-fix-001

Reports are written to build/llamatik-benchmark/ (results.json, results.md, per-task logs).

Current benchmark tasks

ID	Name	Tools exercised
kotlin-simple-fix-001	Fix failing Calculator test	read_file, write_file
kotlin-create-file-002	Create missing StringUtils file	read_file, create_file
kotlin-replace-in-file-003	Fix off-by-one error in Stack	read_file, replace_in_file
kotlin-search-fix-004	Find and fix division bug via search	search_code, read_file, replace_in_file
kotlin-multi-file-005	Fix interface and implementation mismatch	read_file, write_file (two files)
kotlin-add-function-006	Add missing clamp function	search_filenames, read_file, write_file
kotlin-list-files-007	Discover and fix unimplemented method	list_files, replace_in_file
kotlin-git-aware-008	Complete uncommitted work using git tools	git_status, git_diff, read_file, write_file
real-keval-parser-009	[Real project] Keval: parser silently ignores unbalanced )	list_files, read_file, replace_in_file

See docs/benchmarking.md for the full guide, including how to add tasks, the report format, and pass/fail criteria.

---

Free vs Pro

Every feature is free for individual developers. There are no feature gates, no trial periods, and no paywalls.

Pro exists in two forms:

	Individual Pro	Organization Pro
Who it's for	Individual developers who want to support the project	Companies and teams using Llamatik Code at work
Required?	No — purely voluntary	Yes — required for commercial use
What you get	A way to fund continued development; same features as Free	Same features; funds maintenance, new features, and model updates
How to get it	GitHub Sponsors or the Support ♥ button in the plugin	JetBrains Marketplace ($15/month or $150/year per seat)

Individual developers: everything works out of the box, no account needed.
Companies: purchasing an organization subscription is the right way to support the project and keeps it a legitimate procurement line item.

---

🤝 Contributing

Llamatik Code is open-source and welcomes contributions. Please read CONTRIBUTING.md before opening a PR — it documents the dev setup, how to run tests and benchmarks, and a table of known intentional stubs you should be aware of before diving in.

• 🐛 Bug reports → GitHub Issues
• 💡 Feature requests → GitHub Issues
• 📖 Documentation → docs.llamatik.com
• 💬 Community → Discord
• 🤝 Code of Conduct → CODE_OF_CONDUCT.md

Building from source

git clone https://github.com/llamatik/Llamatik-Teammate.git
cd Llamatik-Teammate
./gradlew buildPlugin

The built plugin .zip is written to build/distributions/. Install it via Settings → Plugins → Install Plugin from Disk.

---

📜 License

The source code is licensed under the Apache License 2.0.

The visual assets (logos, icons, and illustrations in assets/ and src/main/resources/icons/) are proprietary — see LICENSE-assets for the terms that apply to them.

See NOTICE for third-party attributions.

---

Built with ❤️ for developers who care about privacy.