Out of scope. This page describes the agent-body MCP gateway from the earlier eleven-organ design. For the MCP tools agent-brain exposes, see the agent-brain CLI reference.

Model Context Protocol (MCP) Integration

The Autonomic ecosystem integrates seamlessly with modern AI coding environments—including Cursor, Claude Desktop, VSCode, Claude Code, and Antigravity—via the Model Context Protocol (MCP).

Autonomic supports Dual-Mode MCP Transport:

1. Standalone Stdio Mode: Individual organs (like agent-brain) run as independent stdio processes spawned directly by the IDE.
2. Unified Gateway Mode: autonomic serve-mcp boots as a single master MCP gateway that dynamically aggregates tool definitions across all running daemons over the internal NATS JetStream bus.

flowchart TD
    subgraph Hosts["Host AI Coding Environments"]
        Cursor["Cursor IDE"]
        Claude["Claude Desktop / Code"]
        Anti["Antigravity / Gemini CLI"]
    end

    subgraph DualMode["Dual-Mode MCP Transport Layer"]
        Gateway["Unified Gateway\n(autonomic serve-mcp)"]
        Direct["Standalone Stdio Proxy\n(agent-brain serve)"]
    end

    Hosts -->|stdio / JSON-RPC| Gateway
    Hosts -->|stdio / JSON-RPC| Direct

    Gateway --> Mesh["NATS JetStream Bus (:4222)"]
    Direct --> Sock["serve.sock\n(Lock-Free Background Daemon)"]

    subgraph Organs["11 Distributed Organs"]
        Mesh --> Body["agent-body (Catalog)"]
        Mesh --> Brain["agent-brain (Routing)"]
        Mesh --> Spine["agent-spine (Workflows)"]
        Mesh --> Muscle["agent-muscle (Execution)"]
        Mesh --> Immune["agent-immune (Security)"]
        Mesh --> Heart["agent-heart (Budget)"]
        Mesh --> Eyes["agent-eyes (Vision)"]
        Mesh --> Mouth["agent-mouth (Ingress)"]
    end

    classDef host fill:#0f172a,stroke:#64748b,stroke-width:1.5px,color:#f8fafc;
    classDef transport fill:#081c2e,stroke:#38bdf8,stroke-width:1.5px,color:#f8fafc;
    classDef organ fill:#04211a,stroke:#10b981,stroke-width:1.5px,color:#f8fafc;
    class Cursor,Claude,Anti host;
    class Gateway,Direct,Mesh,Sock transport;
    class Body,Brain,Spine,Muscle,Immune,Heart,Eyes,Mouth organ;

The route_task Enforcement Gate

When an agent executes in your IDE, agent-brain intercepts the turn using host shell and MCP hooks:

1. Interception: Any attempt to run shell tools or modify files is blocked until route_task is called.
2. Context Synthesis: route_task evaluates the user’s intent and returns a token-bounded (~500 tokens) payload of ranked skills, codebase rules, and verified recipes.
3. Unlocking: Once context is injected, all other tools are unlocked for the agent’s turn.

Editor Configuration & Quick Install

Configure your development environment with one automated command:

# Global setup: automatically detects and configures Cursor, Claude Desktop, and Claude Code
agent-brain install --global

# Target specific IDEs individually:
agent-brain install --cursor
agent-brain install --claude-desktop
agent-brain install --opencode

Manual Configuration Snippet (~/.cursor/mcp.json)

{
  "mcpServers": {
    "agent-brain": {
      "command": "agent-brain",
      "args": ["serve"]
    },
    "autonomic-cluster": {
      "command": "autonomic",
      "args": ["serve-mcp", "--all"]
    }
  }
}

Upstream MCP Server Management

agent-brain acts as an MCP gateway and tool router. External MCP servers can be added, tested, and indexed directly through the CLI:

# Register an upstream MCP server
agent-brain mcp add memory-engine -- /usr/local/bin/custom-mcp-server --port 8080

# List registered servers and active tool counts
agent-brain mcp list

# Test upstream connectivity and verify tool schemas
agent-brain mcp test memory-engine

# Sync tool vectors into SQLite vector memory
agent-brain mcp sync

Skill Packages & Curated Registry

Skills and domain policies are managed via the canonical package command group:

# Canonical package installation
agent-brain package add @agstack
agent-brain package add @supervisor
agent-brain package add owner/repo --ref main

# List installed packages and filesystem paths
agent-brain package list

# Update or remove packages
agent-brain package update
agent-brain package remove coding-policy

# Browse curated ecosystem registry
agent-brain registry list

Universal Upstream Tool Governance (route_to_mcp)

The central router route_to_mcp equips calling agents with high-performance primitives for managing external MCP servers:

1. Pre-Flight Schema Introspection (describe_tool)

Callers can inspect input parameter schemas before dispatching tools, avoiding trial-and-error argument guessing:

{
  "name": "route_to_mcp",
  "arguments": {
    "server": "linear",
    "tool": "save_issue",
    "describe_tool": true
  }
}

2. Multi-Tool Batching (batch)

Executes an array of sequential tool calls within a single persistent child process session, eliminating repetitive process spawn and handshake overhead:

{
  "name": "route_to_mcp",
  "arguments": {
    "server": "linear",
    "batch": [
      { "tool": "save_label", "arguments": { "name": "bug" } },
      { "tool": "save_issue", "arguments": { "title": "Fix crash", "team": "ENG" } }
    ]
  }
}

3. Dynamic Semantic Entity Compaction (compact_entity)

When token limits are constrained, arrays of objects are dynamically compacted to their core identifiers and state fields (id, key, title, name, status, state) before falling back to item omission, maximizing usable context without blowing token budgets.

Token-Bounded Filesystem Tools (agent-brain)

To prevent catastrophic token expansion from unbounded shell commands (ls -la, find, cat), agent-brain provides first-class, token-budgeted filesystem exploration tools:

In-Gateway AST Validation & Self-Correcting Execution (agent-body)

The agent-body gateway automatically mediates shell command execution through AST safety analysis:

1. Automated Gate Fusion: Calls to muscle_execute_bash pass through agent-mouth AST safety validation internally. No separate roundtrip is required.
2. Deterministic Blocking: Destructive commands (e.g. rm -rf /, fork bombs, unvalidated dynamic evaluation) are rejected with exit code 126 and structured feedback detailing the offending AST pattern.
3. Mandatory Self-Correction: Agents are governed to diagnose and correct rejected commands rather than attempting to bypass safety gates with host commands.
4. Resilient Auto-Recovery: Failed organ sessions are cleanly evicted, ensuring that fallback execution does not latch onto subsequent invocations.