lava-mcp

An MCP server that exposes a LAVA instance to agents — letting them query the board farm, submit and manage test jobs, and (in a later phase) open interactive sessions to a board.

It is a thin client over LAVA's REST API (v0.2/v0.3); point it at any LAVA instance.

Install

pip install -e .[dev]

Related MCP server: A2A Client MCP Server

Credentials

The LAVA target (LAVA_URL) is normally pinned to the instance a deployment fronts; the token is resolved per request:

• Hosted (HTTP), pinned — set LAVA_URL on the server to the LAVA instance it serves. Each connecting client then sends only its own X-Lava-Token and acts as its own LAVA user; no X-Lava-Url is needed. The server stores no per-user token.

• Hosted (HTTP), multi-tenant — leave LAVA_URL unset; each client sends both X-Lava-Url and X-Lava-Token, so one server can front many LAVA instances.

• Local (stdio) mode — falls back to LAVA_URL / LAVA_TOKEN in the environment (single user).

Run (stdio, local)

export LAVA_URL=https://lava.example.com
export LAVA_TOKEN=<your-api-token>
lava-mcp                                  # or: lava-mcp --read-only

Launch over stdio from your MCP client (claude_desktop_config.json / Claude Code):

{
  "mcpServers": {
    "lava": {
      "command": "lava-mcp",
      "env": { "LAVA_URL": "https://lava.example.com", "LAVA_TOKEN": "..." }
    }
  }
}

For a hosted server pinned to a LAVA instance, point Claude Code at the HTTP endpoint and pass only your token:

claude mcp add --transport http lava https://mcp.example.com/mcp \
  --header "X-Lava-Token: <your-api-token>"

If the server is multi-tenant (no LAVA_URL set), also pass --header "X-Lava-Url: https://lava.example.com" to choose the instance.

Run as a hosted service (HTTPS via Caddy)

For interactive board sessions the server must be reachable by lab workers, so run it hosted. docker compose brings up the MCP server behind Caddy (automatic HTTPS) and exposes the SSH board-session gateway:

cp .env.example .env      # set LAVA_URL/LAVA_TOKEN/LAVA_MCP_DOMAIN/LAVA_MCP_GATEWAY_HOST
docker compose up -d

• Agents connect to https://$LAVA_MCP_DOMAIN/mcp (streamable-HTTP transport).

• In-job containers dial the SSH gateway at $LAVA_MCP_GATEWAY_HOST:2222.

Or run the HTTP transport directly:

lava-mcp --transport streamable-http --host 0.0.0.0 --port 8000 --gateway

Interactive board sessions (gateway)

In hosted mode with --gateway (or LAVA_MCP_GATEWAY_ENABLED=true), the server runs an in-process SSH rendezvous. open_board_session submits a LAVA job that runs a device-attached container; the container dials out (ssh -R) to the gateway, so no inbound access to the worker is needed. Then:

• run_in_session(session_id, command) runs a command on the board's container (e.g. qdl, fastboot, adb, shell).

• close_board_session(session_id) cancels the job and frees the board.

The container image + test definition live in this repo under interactive/ (published to ghcr.io/mattface/lava-mcp/interactive and fetched from this repo by the lab worker); the parameter contract is in lava_mcp/jobs.py.

Configuration

Tools (v1)

Read/observe: whoami, version, list_devices, get_device, get_device_dictionary, list_device_types, list_workers, list_jobs, get_job, get_job_definition, get_job_logs, get_job_results, get_queue, get_running, get_lab_health, validate_job.

Write (omitted with --read-only): submit_job, cancel_job, resubmit_job.

Interactive board sessions (hosted gateway mode): open_board_session, run_in_session, close_board_session, list_board_sessions.

Test

pytest

Roadmap

• The interactive board sessions gateway is implemented here, along with the container image + test definition the in-job container runs (interactive/, published to ghcr.io/mattface/lava-mcp/interactive).

• Human shell proxy + interactive PTY through the gateway.