Agents & MCP

What an AI agent can do on LabMission

Any agent that speaks the Model Context Protocol can sign in to your workspace, read it, query and stream its data, and draft new integrations, queries and dashboards. Each change it drafts is a proposal that a person approves. Everything an agent does, your team can also do by hand.

Your agent → LabMission MCP server
  • 01get_workspace_flow→14 nodes · 2 surfaces · 0 pending
  • 02list_data_connections→Freezer B, PR-2, Courier tracking
  • 03run_warm_queryfreezer_temps_24h→1,440 rows
  • 04tail_telemetryfreezer-b · 60 s→60 readings, live
  • 05manage_proposalscreate→#128 awaiting review
  • Signed in as agent · rights: Biobank operator (read) + propose
How agents connect

One governed door for every agent

The MCP server is how an agent reaches LabMission. It signs the agent in, applies the rights you gave it, and records every call. Behind it, the agent works with the same workspace your team uses.

ANY MCP CLIENTTHE GOVERNED DOORWHAT THE AGENT CAN DOYOUR WORKSPACEClaudeChatGPTMicrosoft CopilotCursor & IDE agentsYour own agentLabMissionMCP serverOAuth sign-inYour access rights applySecrets stripped from readsSchemas & examples servedEvery call on the recordRead the workspaceget_workspace_flowFind data connectionslist_data_connectionsRun a queryrun_warm_query · ad-hoc KQLTail a live streamtail_telemetryExport historydownload_cold_dataPropose a changemanage_proposalsConnectionsData streamsQueriesSurfacesAgentsWorkbenchesA person reviewsApprove, change or rejectDeployed and recordedSigned, hashed, retainedReads are scoped to the agent's rights. Its only write is a proposal, and a person approves it before anything changes.
What an agent can do

Six things an agent does, and an example of each

Read the workspace

get_workspace_flow

Sees every connection, query, surface and agent, how they connect, and what proposals are pending. Secrets are stripped before it sees anything.

Ask"What feeds the cold-chain dashboard?"

Find data sources

list_data_connections

Lists the instruments, sensors and systems already connected, so it builds on what exists instead of duplicating it.

Ask"Is plate reader PR-2 connected yet?"

Query the data

run_warm_query

Runs your named queries, or writes its own in KQL, over the data your lab is producing.

Ask"How long was Freezer B above -70 °C this week?"

Watch a live stream

tail_telemetry

Follows readings as they arrive, for up to five minutes, filtered to one connection or one surface.

Ask"Watch the incubators while the door test runs."

Pull history

download_cold_data

Exports a time range as CSV or JSONL for analysis, a report or a downstream system.

Ask"Export last quarter's freezer readings."

Propose a change

manage_proposals

Drafts new connections, queries, surfaces, data stores, API sources and access rights as one proposal. This is the agent's only way to change anything.

Ask"Connect PR-2 and put its results on the samples."

Agents also read the API's OpenAPI spec, filtered to what they're allowed to call, plus a schema and worked examples for every kind of node they can propose.

An agent at work

An agent builds a freezer-monitoring integration

A biobank manager asks for alerting on a new ultra-low freezer. The agent does the building and the manager makes the decision.

  1. 1
    You

    Ask in plain language

    "Freezer B's new sensor is online. Alert the custodian when it rises above -70 °C, and show which samples were inside."

  2. 2
    Agent

    Reads what exists

    Reads the workspace, finds the sensor's connection, and tails a minute of readings to learn their shape.

  3. 3
    Agent

    Drafts the change

    Proposes a query for excursions, an alert on that query, and a panel on the storage dashboard, each with its reason.

  4. 4
    You

    Review and sign

    Read the proposal as a diff, adjust the threshold, and approve it. Where your process needs a second approver, it waits for them.

  5. 5
    LabMission

    Deploys and records

    The change goes live. The proposal, the approval and the signature are hashed into the record.

  6. 6
    Your lab

    Excursions reach the right person

    The next excursion opens one impact item per tube in Sample Management, for the custodian to claim and decide.

Assistant · Biobank · Site A
Alert the custodian when Freezer B rises above -70 °C, and show which samples were inside.
I read the workspace and tailed a minute of Freezer B's readings. I've drafted these changes for your review. Nothing is live yet.
Proposal #131Awaiting review
  • + query      freezer_b_excursions
  • + alert      Custodian, on each excursion
  • ~ surface    Storage dashboard (adds 1 panel)
Reason: new freezer commissioned, change request CR-2026-131 · Two-person approval required
Approve & signRequest changes
Sample Management's excursion impact assessment: one item per tube, each to be claimed and decided as acceptable, quarantine or destroy.
What agents build

Integrations and tools a lab asks for

These are the requests agents handle well, because each one is built from parts the platform already has.

Instrument integrations

A connection for a reader or robot, the column mapping for its export, and a query over its results.

Cold-chain monitoring

Sensor connections, excursion queries and alerts that resolve to the samples that were exposed.

Courier and partner feeds

An API source that polls a courier or partner on a schedule, so expected arrivals reach reception.

LIMS and ELN bridges

A connection that receives events from systems you already run and lands them on the same record.

Study dashboards

Surfaces and interfaces over the queries a study team watches, published for the people who need them.

Workbench code

The code for a workbench, such as a label service or a calculation, proposed and deployed like any change.

Access proposals

Draft access rights and role configurations for a new team, for an administrator to accept.

Analysis extracts

Time-bounded exports and published queries for statisticians, data scientists and downstream tools.

Governance

The same rules for agents as for people

ClaudeChatGPTMicrosoft CopilotCursorYour own agents
  • An agent signs in with OAuth and holds only the rights you grant
  • It reads through those rights, and secrets never reach it
  • Its only write is a proposal; a person approves, changes or rejects it
  • Approvals can require a reason, a signature and a second person
  • Every call and every change lands in the same provenance record
  • AI arrives switched off and runs on your own AI service when you turn it on

Watch an agent build on LabMission

We'll connect an agent to a workspace, have it draft an integration, and walk the proposal through review, signature and deployment.