MCP Integration

Feature: F022, F044, F045, F046, F047 | Status: Shipped in v0.12.0

Cognition Engines exposes 14+ tools via the Model Context Protocol for integration with Claude Code, Claude Desktop, OpenClaw, and any MCP-compliant agent.

Endpoint

http://your-server:9991/mcp

Streamable HTTP - handles both POST (tool calls) and GET (SSE events).

Primary Tools (Start Here)

Tool	Description
pre_action (PRIMARY)	All-in-one pre-action check: queries similar past decisions, evaluates guardrails, fetches calibration context, extracts patterns, and optionally records the decision. One call replaces query_decisions + check_action + log_decision.
get_session_context (PRIMARY)	Full cognitive context for session start: agent profile (accuracy, Brier, tendency), relevant decisions, guardrails, calibration by category, overdue reviews, and confirmed patterns. Available in JSON or markdown for system prompt injection.
ready (PRIMARY)	Prioritized cognitive maintenance queue: overdue reviews, calibration drift, stale decisions. Filter by priority, type, category.

Graph Tools (F045)

Tool	Description
link_decisions	Create typed edges between decisions (relates_to, supersedes, depends_on) with optional weight
get_graph	Query subgraph around a decision with configurable depth and edge type filters. Returns nodes with metadata and weighted edges.
get_neighbors	Lightweight neighbor query - direct connections for a decision with edge types and weights

Granular Tools (Fine-Grained Control)

Tool	Description
query_decisions	Search past decisions with optional bridge_side (prefer pre_action)
check_action	Validate against guardrails (prefer pre_action)
log_decision	Record a decision manually — last resort when pre_action wasn’t used
review_outcome	Record what actually happened
get_stats	Calibration statistics (prefer get_session_context)
get_decision	Full decision details by ID
get_reason_stats	Which reason types predict success
update_decision	Update tags/pattern on existing decisions
record_thought	Capture chain-of-thought reasoning steps

Recommended Workflow

Session start    → get_session_context       (load cognitive context)
       ↓
Maintenance      → ready                     (check for overdue reviews, drift)
       ↓
Decision point   → pre_action                (query + guardrails + record)
                   (auto_record: true)        → returns decisionId
       ↓
During work      → record_thought            (capture reasoning)
                   (decision_id: from above)  → thoughts attach in real-time
       ↓
After work       → update_decision           (finalize decision text)
                   (id: decisionId)
       ↓
Link related     → link_decisions            (explicit relationships)
       ↓
Later            → review_outcome            (record success/failure)

Multi-Agent Isolation

When multiple agents share a single MCP connection (e.g. Claude Code sub-agents), pass agent_id to isolate deliberation streams:

pre_action(agent_id: "planner", ...)           → decisionId: "abc123"
record_thought(agent_id: "planner", decision_id: "abc123", text: "...")
update_decision(id: "abc123", ...)

Each agent’s thoughts are tracked via composite keys (agent:{id}:decision:{id}), preventing cross-contamination.

When to Use log_decision

log_decision is a last resort for recording decisions without prior context. Use it only when:

• The agent didn’t call pre_action (legacy or spontaneous decision)
• You need full manual control over all fields
• No deliberation tracking is needed

For the standard flow, pre_action(auto_record: true) + record_thought + update_decision is always preferred.

Claude Code CLI

Add to your project’s .mcp.json (or global ~/.claude.json):

{
  "mcpServers": {
    "decisions": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "http://your-server:9991/mcp",
        "--allow-http",
        "--header",
        "Authorization: Bearer YOUR_CSTP_TOKEN"
      ]
    }
  }
}

On Windows, use cmd as the command:

{
  "mcpServers": {
    "decisions": {
      "command": "cmd",
      "args": [
        "/c", "npx", "mcp-remote@latest",
        "http://your-server:9991/mcp",
        "--allow-http",
        "--header",
        "Authorization: Bearer YOUR_CSTP_TOKEN"
      ]
    }
  }
}

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "decisions": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "http://your-server:9991/mcp",
        "--allow-http",
        "--header",
        "Authorization: Bearer YOUR_CSTP_TOKEN"
      ]
    }
  }
}

OpenClaw Configuration

Add to your OpenClaw gateway config:

mcp:
  servers:
    cognition-engines:
      url: http://your-server:9991/mcp
      headers:
        Authorization: "Bearer your-token"

Stdio Transport

python -m a2a.mcp_server

Example: Pre-Action Tool

{
  "name": "pre_action",
  "arguments": {
    "action": {
      "description": "Refactor auth module to use JWT",
      "category": "architecture",
      "stakes": "high",
      "confidence": 0.82
    },
    "reasons": [
      {"type": "analysis", "text": "Stateless auth scales better for microservices"},
      {"type": "pattern", "text": "Team successfully used JWT in 2 other services"}
    ],
    "tags": ["auth", "jwt", "refactor"],
    "pattern": "Stateless auth scales better than session-based"
  }
}

Returns: allowed status, relevant past decisions, guardrail results, calibration context, confirmed patterns, and optionally a recorded decision_id.

Example: Session Context Tool

{
  "name": "get_session_context",
  "arguments": {
    "task_description": "Build authentication service for user API",
    "format": "markdown"
  }
}

Returns: markdown-formatted cognitive context ready for system prompt injection, including agent profile, relevant decisions, guardrails, calibration by category, and confirmed patterns.

FORGE Plugin for Claude Code

FORGE is a Claude Code plugin that implements the full decision loop as a protocol:

Fetch → Orient → Resolve → Go → Extract

Phase	What happens	Tool used
Fetch	Load cognitive context at session start	get_session_context
Orient	Check guardrails before non-trivial decisions	pre_action
Resolve	Stream micro-thought deliberation (10+ signals/decision)	record_thought
Go	Execute with confidence, gate HIGH/CRITICAL on confirmation	update_decision
Extract	Review outcomes, distill patterns	review_outcome

Installation

claude --plugin-dir ./path-to-forge-plugin

Components

• Skill (skills/forge-protocol/SKILL.md) - Protocol rules loaded contextually
• Hooks (hooks/hooks.json) - SessionStart injects protocol, PreToolUse gates risky ops, Stop enforces reflection
• MCP (.mcp.json) - Connects to the decisions server
• Commands - /forge:status dashboard, /forge:reflect batch review

→ FORGE Repository

vs JSON-RPC

MCP adds session management overhead. For stateless workflows, the JSON-RPC API via cstp.py is simpler. Use MCP when your agent platform expects it.