EVC Mesh MCP Server

Model Context Protocol (MCP) server for EVC Mesh — a task management platform for coordinating humans and AI agents.

Connects AI agents (Claude Code, Cursor, Cline, OpenClaw, etc.) to EVC Mesh via MCP tools for task management, persistent memory, event publishing, and multi-agent coordination.

This is the actively developed copy. evc-mesh also ships an MCP server (./cmd/mcp, same internal/mcp tool set) that it builds and deploys itself — the two exist because Go's internal/ visibility rules mean one repo can't import the other's package, not because they're meant to diverge. New tools and fixes land here first.

Prerequisites

• Go 1.22+

• Running EVC Mesh instance

• Agent registered in Mesh with an API key (agk_...)

Related MCP server: slm-mesh

Installation

go install github.com/entire-vc/evc-mesh-mcp@latest

Or build from source:

git clone https://github.com/entire-vc/evc-mesh-mcp.git
cd evc-mesh-mcp
go build -o evc-mesh-mcp .

Docker

docker run -i --rm \
  -e MESH_API_URL \
  -e MESH_AGENT_KEY \
  ghcr.io/entire-vc/evc-mesh-mcp

-i is required — the server speaks MCP over stdio, and Docker only wires up stdin when the container runs interactively. Add -e MESH_MCP_PROFILE=core to switch profiles (see Tool Profiles below). The image is published for linux/amd64 and linux/arm64 from Dockerfile in this repo on every tagged release (docs/RELEASING.md).

Tool Profiles

The MCP server supports two profiles to optimize context window usage:

Set via MESH_MCP_PROFILE environment variable. Default: full.

Configuration

Running without credentials

In stdio mode the server also starts when MESH_AGENT_KEY is not set. It then answers initialize and tools/list as usual, and every tool call returns instructions for setting MESH_API_URL and MESH_AGENT_KEY. This lets MCP clients and catalogs inspect the tool list before you have a key. If a key is set but authentication fails at startup (API unreachable, key rejected), the server keeps running: tools are listed, and each call retries authentication and returns the reason until it succeeds.

Tool annotations

Every tool declares the MCP hints readOnlyHint, destructiveHint, idempotentHint and openWorldHint, so clients can tell read-only tools (get_*, list_*, recall, search_docs, …) from ones that change or remove data (update_*, move_task, forget, …).

Client metrics

Each initialize is logged with the client's clientInfo.name and version, and counted in the Prometheus metric mesh_mcp_initialize_total{client,profile} (exposed on /metrics in SSE mode; client names are normalised and capped).

Claude Code (stdio mode)

Add to your project's .mcp.json:

{
  "mcpServers": {
    "evc-mesh": {
      "command": "evc-mesh-mcp",
      "env": {
        "MESH_API_URL": "https://your-mesh-instance.example.com",
        "MESH_AGENT_KEY": "agk_your-workspace_your-key",
        "MESH_MCP_PROFILE": "core"
      }
    }
  }
}

Cursor

Add to Cursor MCP settings (Settings → MCP Servers):

{
  "evc-mesh": {
    "command": "evc-mesh-mcp",
    "env": {
      "MESH_API_URL": "https://your-mesh-instance.example.com",
      "MESH_AGENT_KEY": "agk_your-workspace_your-key",
      "MESH_MCP_PROFILE": "core"
    }
  }
}

SSE Mode (multi-agent, shared server)

For connecting multiple agents through a shared MCP endpoint:

MESH_API_URL=https://your-mesh-instance.example.com \
MESH_MCP_PORT=8081 \
evc-mesh-mcp --transport sse

SSE mode serves two profiles simultaneously on different paths:

The same process also serves the Streamable HTTP transport (stateless, one agent key per request, sent in the Authorization: Bearer or X-Agent-Key header — the query parameter is refused there):

Authentication per connection via:

• Authorization: Bearer agk_... header

• X-Agent-Key: agk_... header

• ?agent_key=agk_... query parameter

Agent Context Protocol (ACP)

At session start, follow these 5 steps in order:

1. heartbeat(status="online")              → register as alive
2. get_project_knowledge(project_id)       → load accumulated decisions & conventions
3. get_my_rules(project_id)                → understand constraints
4. get_context(project_id)                 → see recent activity + project knowledge
5. get_my_tasks()                          → check assigned work

At session end:

publish_event(type="summary", memory={persist: true})  → broadcast + persist
session_report(model, tokens_in, tokens_out)           → report metrics

MCP Tools — Core Profile (25)

ACP & Identity

Task Management

Communication

Memory

What recall guarantees about its result

limit is a hard bound. The response never contains more than limit items, and total always equals the number of items actually returned. Nothing is added to the page after it has been sized — not pinned rows, not graph-expanded neighbours.

Rows that fail scope/tags/tags_any are dropped, never returned unmarked. This holds regardless of how a row reached the result: ordinary retrieval, pinning, or graph expansion. A pinned row is exempt from ranking, not from eligibility — "pinned" means "do not let ranking bury this", not "show this to a caller who asked for a different scope".

Graph neighbours are marked and bounded. With RECALL_GRAPH_ENABLED=true, recall also runs a knowledge-graph expansion and folds in hop > 0 neighbours, each carrying graph_boost: true and provenance: via:graph. They occupy at most limit/4 of the page (at least 1 when limit >= 2, none when limit < 2) and take its tail slots, displacing the weakest retrieval hits rather than being appended on top. When expansion returns nothing usable, the page is exactly the base result — the reserve is a ceiling, not a quota. graph_boost_count reports how many slots were actually spent.

The reserve exists because base hits carry score (RRF across the retrieval arms) and neighbours carry composite_score from a separate traversal — different fields on different scales. Sorting the union on a common key does not balance them; in practice every observed neighbour ranks below every base hit, so a naive merge-sort would silently disable graph boost. The reserve makes that trade explicit and tunable.

Presets never overrule you. recall classifies the query and may apply a profile (e.g. multi-session widens the page). A profile only fills in parameters you did not supply; an explicit limit always wins.

Utility

MCP Tools — Full Profile (adds 38 more, 63 total)

Additional Task Tools

Comments & Artifacts

Downloading an artifact

Downloading an artifact is two GETs. Step 1: GET /api/v1/artifacts//download with header X-Agent-Key: -> 200 JSON {"url": ""}. Step 2: GET that url with NO headers -> 200, the file bytes. Pitfalls: on step 1 only X-Agent-Key is accepted (X-API-Key and Authorization: Bearer give 401); on step 2 any extra header, Authorization in particular, breaks the presigned signature (400). The artifact's download_path is step 1's path. Never fetch browser_only_url with an agent key: it is a human page and answers 401 by design.

Event Bus

Agent & Team

Governance & Config

Recurring Tasks

Documents & Knowledge

Architecture

AI Agent (Claude Code / Cursor / Cline / OpenClaw)
    ↕ MCP (stdio or SSE)
EVC Mesh MCP Server (core or full profile)
    ↕ REST API (HTTP)
EVC Mesh API Server
    ↕
PostgreSQL / Redis / NATS / S3

The MCP server is a lightweight proxy — it translates MCP tool calls into REST API requests. No direct database access needed.

Running the shared HTTP server

To serve several agents from one process, run the server in SSE mode next to your Mesh API (the same image works: docker run -e MESH_MCP_TRANSPORT=sse -e MESH_API_URL=... -p 8081:8081 ghcr.io/entire-vc/evc-mesh-mcp) and put it behind your reverse proxy. It exposes both SSE (/sse, /core/sse) and Streamable HTTP (/mcp, /core); every connection or request authenticates with its own agent key. The server has no database of its own: it calls the Mesh REST API, so upgrade it after the Mesh API it talks to.

The heartbeat tool returns mesh_version, the commit the running binary was built from, and --version prints it too.

Related

• evc-mesh — Core platform (API + Web UI)

• evc-mesh-openclaw-skill — OpenClaw skill (bash scripts)

License

MIT