ix-openclaw-plugin

An OpenClaw plugin that turns your agent into a graph-reasoning engineering agent using Ix Memory as its structured memory backend.

OpenClaw + Ix = reasoning engine + persistent code knowledge graph. Skills are cognitive abstractions (not CLI wrappers) that minimize token usage and maximize accuracy.

Installation

openclaw plugins install ix-infrastructure/ix-openclaw-plugin && openclaw gateway restart

Requirements

Ix Memory backend

Ix Memory must be installed and running:

ix status   # should return ok
ix map      # index the codebase if not already done

Agent model — gateway-based only

This plugin's tools and skills run through the OpenClaw gateway. The OpenClaw Codex sandbox harness does not support plugin tools and will not work. Use any gateway-based model instead:

Provider	Setup
Anthropic (recommended)	Get a key at console.anthropic.com
OpenAI platform	Get a key at platform.openai.com (not a ChatGPT account)
Groq (free tier)	Get a key at console.groq.com
Google Gemini (free tier)	Get a key at aistudio.google.com

Wiring an agent

1. Add your API key to the agent's auth-profiles file:

# Find your agent's auth file
cat ~/.openclaw/agents/<your-agent-id>/agent/auth-profiles.json

Add an entry for your provider (example for Anthropic):

{
  "version": 1,
  "profiles": {
    "anthropic:manual": {
      "type": "token",
      "provider": "anthropic",
      "token": "sk-ant-..."
    }
  }
}

2. Set the model in ~/.openclaw/openclaw.json:

Find your agent in agents.list and set the model:

{
  "agents": {
    "list": [
      {
        "id": "your-agent-id",
        "workspace": "/path/to/your/project",
        "model": "anthropic/claude-sonnet-4-6"
      }
    ]
  }
}

Supported model strings: anthropic/claude-sonnet-4-6, anthropic/claude-opus-4-7, openai/gpt-4o, groq/llama-3.3-70b-versatile, google/gemini-2.0-flash.

3. Restart the gateway:

openclaw gateway restart

Ix Pro is optional. All skills and hooks work with basic Ix. Pro adds the session briefing (goals, bugs, decisions) that the plugin prepends to prompts.

Skills

High-level cognitive skills — each one infers intent, orchestrates multiple graph queries, and synthesizes output. None are CLI aliases.

Skill	What it does	Key rule
/ix-understand [target]	Build a mental model of a system or the whole repo	Graph only — no source reads
/ix-investigate <symbol>	Deep dive: what it is, how it connects, execution path	Graph first, one symbol read max
/ix-impact <target>	Change risk: blast radius, affected systems, test targets	Depth scales with risk level
/ix-plan <targets...>	Risk-ordered implementation plan for a set of changes	Parallel impact, finds shared dependents
/ix-debug <symptom>	Root cause analysis from symptom to candidates	Targeted reads at suspects only
/ix-architecture [scope]	Design health: coupling, smells, hotspots	Graph only — never reads source
/ix-docs <target> [--full] [--style narrative|reference|hybrid] [--split] [--single-doc] [--out <path>]	Generate narrative-first system documentation with a selective reference layer	Default is onboarding-focused; --full --style hybrid gives the deepest coverage

All skills auto-disable when ix is not available (via requires.bins gating).

Agents

Autonomous multi-step agents for complex tasks:

Agent	Purpose
ix-explorer	General-purpose graph exploration, open-ended questions
ix-system-explorer	Full architectural model of a codebase or region
ix-bug-investigator	Autonomous investigation from symptom to root cause candidates
ix-safe-refactor-planner	Blast radius + safe change sequencing for refactors
ix-architecture-auditor	Full structural health report with ranked improvements

Automatic hooks

All of them are typed plugin hooks (api.on(...) in plugins/ix-plugin.ts); the plugin ships no folder hooks.

Event	Tools	Effect
before_prompt_build	—	Prepends the session briefing (goals, bugs, decisions) once per 10 min per workspace, as text capped at 2,000 characters — requires Ix Pro and allowConversationAccess (see Configuration)
before_tool_call	edit, write, apply_patch	Runs ix-decide on the files being written (risk and dependent counts from ix impact): REVIEW asks you to approve the call (allow once or deny), ALLOW is silent. Nothing is blocked outright. A file Ix rates medium or high risk, or with 5+ dependents, gets a REVIEW; a critical file or one with 20+ dependents gets a REVIEW marked high risk (critical severity, reason shown). Only a risky answer from Ix asks: a low-risk file, a new file, or an edit Ix cannot assess (not installed, backend down, project not mapped), is ALLOW
after_tool_call	edit, write, apply_patch	Requests the guarded root map (below)
session_end	—	Requests the guarded root map for the agent workspace

All hooks bail silently if ix is not in PATH or the backend is unreachable. Search calls (read, exec) get no Ix context: an OpenClaw before_tool_call handler can block, rewrite or ask for approval, but cannot add context for the model.

Guarded root map. Automatic refresh runs ix map <git root> --silent in the background (with IX_AUTO_MAP=1) only when the project is a git repo whose root is not $HOME, ix status reports the project is already mapped, and no automatic map started for that root in the last 5 minutes (IX_MAP_DEBOUNCE_SECONDS). It never creates a workspace — run ix map yourself once per project. Debounce stamps live in ${XDG_STATE_HOME:-~/.local/state}/ix-openclaw-plugin/.

Configuration

Plugin config in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "entries": {
      "ix-memory": {
        "enabled": true,
        "hooks": { "allowConversationAccess": true }
      }
    }
  }
}

allowConversationAccess is needed only for the session briefing: OpenClaw drops a non-bundled plugin's before_prompt_build hook without it (openclaw plugins inspect ix-memory --runtime then reports the hook as blocked). Tools and the edit hooks work either way.

Uninstall

openclaw plugins uninstall ix-memory

License

Apache License 2.0