Serve a Lab to an agent

Let any MCP agent use a Lab’s tools, with every action checked, approved and recorded by the same rules as your evaluations.

Evaluations test what an agent does in a Lab. Serving a Lab lets an agent use it outside an evaluation, with the same safeguards. Every tool call goes through the gateway: it is checked against your rules, held for approval if needed, and recorded, before it reaches the Lab. The same rule file decides the same way in an evaluation and when served.

WarningPrototype

Serving has been tested with simulated and mock Labs only. It has not been used with real instruments, and it is not a substitute for facility safety controls.

Serve a Lab

Install the optional extra, then serve a registered Lab over MCP’s stdio transport:

uv pip install --python .venv/bin/python '.[serve]'
inspect-labs serve --lab liquid-handler \
  --rules rules.json --approvals approvals.json \
  --stop-file .research/STOP --lab-log .research/session.json

Option	What it does
--lab	A registered Lab (see inspect-labs list)
--rules	Action rules as JSON. Without it, the default rules allow reads and reversible actions and hold irreversible and external actions for approval.
--approvals	Actions a person approved in advance, as a JSON list of {"tool": ..., "arguments": {...}}. A held action runs only if it matches one exactly.
--stop-file	Stop the session when this file appears. Every later action is refused, and the file’s text is recorded as the reason.
--lab-log	Where to write the session’s lab log when the agent disconnects. It is never overwritten.

Point any MCP agent at the command. For example, in an MCP client configuration:

{
  "mcpServers": {
    "lab": {
      "command": "inspect-labs",
      "args": ["serve", "--lab", "liquid-handler", "--lab-log", "session.json"]
    }
  }
}

The agent sees the Lab’s tools. A refused action comes back as an error with the reason, and never reaches the Lab. The agent cannot stop or reconfigure the gateway.

After the session

The session’s lab log records every action, the decision, any approval and whether it ran, then what the Lab reported. It is hash-chained like an evaluation’s lab log, and includes monitor flags such as refused actions.

inspect-labs replay-rules .research/session.json --rules new-rules.json

Replaying rules works the same on a session’s lab log as on an evaluation’s: it shows what the new rules would have decided, without running anything.

From Python

from inspect_labs import DEFAULT_RULES
from inspect_labs.gateway import ApprovedAction, approved_actions
from inspect_labs.serve import LabSession, serve_over_stdio

session = LabSession(
    my_lab,
    DEFAULT_RULES,
    approver=approved_actions([ApprovedAction(tool="drop_tip", arguments={})]),
)
# session.stop("reason") refuses every later action.