MCP Integration¶
The Model Context Protocol (MCP) integration lets MCP-native agents (Codex, Cursor, and others) communicate with Cognibrain directly through the standardized MCP tool interface.
When to Use MCP¶
Use MCP when your agent host natively supports the Model Context Protocol. For agents that can only run shell commands, use the Harness Lifecycle instead.
| Agent Host | Recommended Surface |
|---|---|
| OpenAI Codex | MCP |
| Cursor | MCP |
| Custom MCP host | MCP |
| CI/CD pipelines | Harness CLI |
| Git hooks | Harness CLI |
| Non-MCP agents | Harness CLI or SDK |
Setup¶
Cursor¶
The init command automatically generates Cursor MCP configuration:
To regenerate:
Generated .cursor/mcp.json
Codex¶
Manual MCP Server Start¶
For custom MCP hosts, start the MCP server directly:
This starts an MCP stdio server that any MCP-compatible client can connect to.
Available MCP Tools¶
The MCP server exposes these tool classes:
Context Pack¶
Retrieve compact memory context for a task before the agent begins work.
Tool: cognibrain_context
Input: { "task": "prepare the release patch", "tokenBudget": 1200 }
Output: { "context": "...", "memoriesUsed": 5 }
Coding Context¶
Retrieve codebase-specific corrections and conventions relevant to files being edited.
Tool: cognibrain_coding_context
Input: { "task": "fix auth refresh", "files": ["src/auth.ts"] }
Output: { "corrections": [...], "conventions": [...] }
Action Guard¶
Check a planned action against known guards before executing.
Tool: cognibrain_guard
Input: { "action": "edit src/api/server.ts" }
Output: { "allowed": true, "warnings": [] }
Memory Add¶
Store a durable correction, decision, or fact.
Tool: cognibrain_memory_add
Input: { "content": "Use JWT RS256 for auth tokens", "scope": "repo" }
Output: { "id": "mem_abc123", "stored": true }
Patch Evidence¶
Record files changed, commands run, and memories used during a task.
Tool: cognibrain_patch_evidence
Input: { "task": "auth fix", "files": ["src/auth.ts"], "commands": ["npm test"] }
Output: { "evidenceId": "ev_xyz789" }
Maintenance¶
Inspect health and trigger dream-cycle maintenance.
Tool: cognibrain_health
Input: {}
Output: { "status": "healthy", "memoryCount": 42, "staleCandidates": 2 }
Authentication¶
For local development with solo-dev profile, no authentication is needed.
For daemon-backed deployments with auth enabled:
{
"mcpServers": {
"cognibrain": {
"command": "npx",
"args": ["cognibrain", "mcp", "serve"],
"env": {
"MEMORY_API_KEY": "${env:MEMORY_API_KEY}"
}
}
}
}
MCP vs CLI Lifecycle¶
Both MCP and CLI lifecycle commands hit the same memory engine:
graph TB
subgraph "MCP Path"
Agent1[MCP Agent] --> MCPServer[MCP Server]
end
subgraph "CLI Path"
Agent2[Shell Agent] --> CLI[Harness CLI]
end
MCPServer --> Core[Memory Core]
CLI --> Core
Core --> Store[(Storage)] Choose based on what your agent host supports. The memory behavior is identical regardless of surface.
Troubleshooting¶
MCP server not responding
Ensure the Cognibrain daemon is running: