Manthan MCP server

Exposes the Manthan substrate as an MCP (Model Context Protocol) tool surface.
External AI agents — *Claude Desktop, Cursor, Codex CLI, Copilot's MCP
client, the maestro-platform foundry* — can read and write Manthan
artefacts through the same standard tool-call protocol.

Tools (9)

Name	Read/Write	Purpose
`manthan_artefact_create`	W	Create idea / critique / synthesis / prototype / opportunity / message; lineage parent + relation optional. Auto-tagged `via-mcp` + `source:<client>`.
`manthan_artefact_recent`	R	Most-recent artefacts; filter by kind, capped at 200.
`manthan_artefact_get`	R	One artefact by UUID.
`manthan_agent_list`	R	Personas registered in the substrate.
`manthan_herald_status`	R	Live status of all 3 HERALD nodes (sigil, uptime, peer/session, canary period, tree size). One call.
`manthan_herald_canary`	R	Specific node's signed canary. Re-signed every call.
`manthan_herald_charter`	R	Specific node's signed Charter (operator legal posture).
`manthan_herald_entries`	R	Recent CT-log entries, minimum profile.
`manthan_readyz`	R	Single-call stack health.

Run

stdio (Claude Desktop, Cursor, Codex CLI)

cd manthan/mcp-server
npm install --no-audit --no-fund
npm run start:stdio

http/SSE (headless agents, maestro-platform foundry, automation)

MANTHAN_MCP_TRANSPORT=http MANTHAN_MCP_PORT=47100 npm run start:http
# then: GET local-service/sse

Client config

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json
(macOS) or the equivalent on your platform:

{
  "mcpServers": {
    "manthan": {
      "command": "node",
      "args": [
        "--experimental-strip-types",
        "--no-warnings",
        "manthan/mcp-server/src/index.ts"
      ],
      "env": {
        "MANTHAN_API_URL": "local-service",
        "MANTHAN_SUBSTRATE_URL": "local-service"
      }
    }
  }
}

Restart Claude Desktop. The 9 tools appear under the 🔌 menu and Claude can
read + write Manthan during conversations.

Cursor

Cursor uses MCP via its config UI (Settings → Features → MCP). Same shape
as Claude Desktop — point it at this server's stdio entrypoint with the
two env vars.

Codex CLI

Codex's MCP support is via the mcp block in its workspace config. Same
stdio command + env.

Copilot

GitHub Copilot's MCP integration is rolling out (late 2025/early 2026).
When you wire it, point at the http/SSE transport rather than stdio:
MANTHAN_MCP_TRANSPORT=http, then SSE URL http://manthan-host:47100/sse.

maestro-platform foundry

Run the http/SSE variant inside the same network as the foundry. The
swarm agents (Codex / Gemini / DeepSeek / Qwen) connect via their MCP
client wrappers and use the same 9 tools. Promoted artefacts (per a
Patina threshold or operator marker) are exported from Manthan back into
maestro by P-106 — separate ticket, not in this server.

Provenance

Every artefact created via MCP is auto-tagged:

via-mcp (always)
source:<value> (from the source arg the client passes; absent → source:unknown)

Plus a metadata blob: { via: "mcp", source, submitted_at }. So the
viewer can distinguish "this came from a human's Claude session" from
"this came from a long-running maestro agent" from "this came from the
internal runtime baseline."

Smoke test (against a live Manthan stack)

# Stack must be up — see infra/compose/README.md and the root docker-compose.yml.
echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test","version":"0.0.1"}}}
{"jsonrpc":"2.0","method":"notifications/initialized"}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"manthan_readyz","arguments":{}}}' \
| MANTHAN_API_URL=local-service \
  node --experimental-strip-types --no-warnings src/index.ts

You should see ready: true and all four downstreams report ok.

What's not here yet

Authentication / per-client identity. Today, anyone who can reach the

stdio process or the http port can call any tool. For local-only stdio
this is fine; for the http variant in production, put a reverse proxy
with auth in front, or run inside a Tailscale ACL'd network.

Substrate write attribution. author_agent_id is optional; if the

client doesn't pass one, the artefact is unattributed. Future work:
per-client substrate agent so MCP writes land under e.g. "claude-desktop"
or "maestro-codex" identities.

Direct HERALD envelope sending. Today MCP writes go to the Postgres

substrate (idea-layer). Writing into the HERALD CT log (verifiable
layer) requires the agent SDK with a sigil + key — separate
scope, P-103.5 / P-104 territory.

Patina-aware filtering on reads. P-104 lands a per-sigil reputation

score; once it does, manthan_artefact_recent should grow a
min_patina filter so consumers can ask for "only the high-trust
recent artefacts."

Manthan MCP server