MCP

Drive your Open Culture instruments — Habitat fluidics and MaxWell MEA recording rigs — from Claude (or any Model Context Protocol-capable assistant) in plain language. Open Culture hosts one MCP server in the cloud; connect it once and Claude can read status, run approval-gated operations, and answer questions about your rigs.

Unlike the REST API — which exposes every endpoint of a single device — the MCP surface is compact, high-level, and fleet-aware: each tool addresses a specific rig by its rig_id.

Connect Claude to Open Culture

You'll need

An Open Culture account — the same login you use for the web app.

Coming soon: one-click connect

We're submitting the OpenCulture connector to Anthropic's directory. Once it's listed, you'll be able to add it straight from Claude's built-in connector list — no custom URL needed. Until then, use the steps below.

Via claude.ai

In claude.ai, open Settings → Connectors.
Choose Add custom connector.
For URL, enter https://app.openculture.ai/mcp, and give it a name like OpenCulture.
Claude redirects you to the Open Culture sign-in. Log in with your Open Culture account.
Approve the consent screen. The connector should now show as Connected.
Open a new chat and enable the connector for that conversation (the tools / connector toggle).

That's it — ask Claude something like "Is rig MistyForest ready, and what's its CO₂ right now?" and it will call the tools for you.

How approvals work

Nothing runs without your approval

Claude can read freely, but anything that actuates hardware (a pump move, starting a recording, powering a rig on or off) is recorded as a pending request a human must approve — in the web app's Pending Approvals panel, or via the one-click link Claude gives you. An agent can't approve its own actions. The one exception is the emergency stop (panic button), which runs immediately by design.

Every domain follows the same shape: reads return right away (or as a quick background job you poll); operations are proposed for approval; jobs are tracked with <domain>_list_jobs / <domain>_get_job_status; and approvals go through <domain>_approve_job / <domain>_approve_pending (usually done in the web app).

Where to approve

When Claude proposes an action it gives you a one-click approve link — the quickest path. You can also review everything waiting for you in the Pending Approvals panel of the Open Culture web app, which lists each request with its rig, operation, and parameters to Approve or Decline. Approval always happens in the web app, never inside the chat.

Example prompts

Once connected, just ask — Claude picks the right tools. A few to try (swap in your own rig names):

"Is rig MistyForest ready, and what's its CO₂ and temperature right now?"
"Has CO₂ been stable on MistyForest over the last 24 hours?"
"What have the pumps on MistyForest been doing in the last hour — any failures?"
"Prime the reagent lines on MistyForest." — Claude proposes it; you approve.
"Is the MaxWell rig mea-1 ready to record — server up, disk free?"
"Start a 5-minute recording on mea-1." — proposed for your approval.
"What claims do we have about cortical organoids?"

Habitat — fluidics

Read status & telemetry

Tool	What it does
`habitat_check_system_ready`	Is the rig ready to actuate? `ready` plus blocking reasons (drivers not ready, emergency active, faulted pump, stale heartbeat).
`habitat_get_rig_status`	Cached status descriptor — device/API health, safety state, pump count/faults, pending jobs.
`habitat_get_pump_status`	Per-pump snapshot and which pumps are faulted.
`habitat_get_pump_diagnostics`	Live per-pump diagnostic (encoder / registers), run as a background read job.
`habitat_get_pump_activity`	Action log over a window — what the pumps did (dispenses, primes, feeds, failures), with `level` / `action` / `failed_only` filters.
`habitat_get_sensor_latest` · `habitat_get_sensor_history` · `habitat_get_sensor_stats`	Latest / windowed / summarized CO₂ · temperature · humidity.

Run operations (human-approved)

Tool	What it does
`habitat_run_operation`	Propose a vetted fluidics op — `dispense`, `prime`, `feed`, `wash`, `perfuse`, `sterilize`, pump / valve / peristaltic primitives, and more (`habitat_list_operations` returns the live set).
`habitat_run_experiment`	Author a bespoke plan by chatting with the cloud authoring agent; it validates against a mock rig, then hands it to you to approve.
`habitat_emergency_stop`	Panic button — halts every pump immediately, no approval wait.
`habitat_ask_educator`	Ask a conceptual question about operations, errors, or microfluidics / cell-culture concepts. Read-only.

MaxWell MEA — electrophysiology

Read status

Tool	What it does
`ephys_check_system_ready`	Is the MaxWell rig ready to actuate (mxwserver listening, not busy, disk free, no emergency)? A necessary-but-not-sufficient cached check.
`ephys_get_device_info`	Device identity — `device_kind`, version, `num_wells`, `sampling_rate_hz`, `chip_id`, free disk.
`ephys_get_rig_status`	Cached status descriptor (device kind, chip id, well count, disk free, mxwserver up/down).
`ephys_get_mxwserver_status`	Is the MaxWell server listening, and its version.
`ephys_get_smartplug_status`	The rig's mains-power (smart plug) state.

Run operations (human-approved)

Tool	What it does
`ephys_run_operation`	Propose a canned rig op — `start_recording`, `activity_scan`, `select_electrodes`, `system_on`, `system_off`, `emergency_clear` (`ephys_list_operations` returns the live set).
`ephys_run_experiment`	Author a bespoke MaxWell experiment with the cloud authoring agent; it dummy-validates the script, then hands it to you to approve.
`ephys_emergency_stop`	Panic button — stops acquisition/stimulation and cuts the rig's mains power immediately.
`ephys_ask_educator`	Ask a conceptual question about recordings, activity scans, stimulation, or MEA hardware. Read-only.

Recurring schedules

Tool	What it does
`ephys_create_schedule`	Arm a recurring canned op — e.g. record 5 min every hour for 24 h. Bounded by an end time and/or max occurrences. Because this arms unattended future actuation, it requires a standing delegation you grant ahead of time (`ocs approval delegate --hardware ephys --tier schedule --mode always`) — you can't self-authorize recurring runs in the moment.
`ephys_list_schedules` · `ephys_cancel_schedule`	List your recurring schedules; cancel one (revokes its grant so no further occurrences fire).

Delegate deliberately

A standing --mode always delegation lets an agent actuate hardware unattended for the life of the grant. Grant it only when you mean to, scope it to the specific hardware and window, and revoke it when the run is done (ephys_cancel_schedule stops an individual schedule).

Analysis & data

Beyond driving instruments, the connector exposes tools for analyzing recordings and querying what you know — aimed at data analysis, not hardware control. These are read-oriented; committing a claim requires knowledge-graph writes to be enabled for your tenant.

Spike analysis (SpikeLab)

Tool	What it does
`spikelab_list_tools`	List the deterministic SpikeLab analysis primitives you can run.
`spikelab_run`, `spikelab_result`	Run a primitive (non-blocking → `job_id`) and poll for its result. Pass a stable `workspace_id` to accumulate state so steps (load → curate → compute) build on each other.
`spikelab_create_thread`, `spikelab_send`	Open a SpikeLab session — `educator` (Q&A), `sorter` (spike sorting), or `analyzer` (pipelines + figures) — and converse with it.

Ask your data

Tool	What it does
`analysis_create_thread`, `analysis_send`	Open a general analysis session and ask it to run an analysis or produce a figure.
`graph_ask`	Ask a natural-language question about your knowledge graph; a graph-librarian agent answers, grounded on the graph, with citations.
`search_graph`	Free-text search across claims, recordings, analysis workspaces, samples, and more.

Recordings, claims & provenance

Tool	What it does
`find_recordings`	Find recordings by modality, probe, origin, or reference — your own plus the shared global pool.
`recording_windows`	The timing windows a recording carries — task/trial, stimulation, drug exposure, treatment, media, environment.
`recording_treatment_history`	How a culture was fed and treated leading up to a recording (media-state chain + fluidics schedule).
`find_claims`, `find_claims_by_sample`, `get_claim`	Browse and read hypotheses / insights, optionally scoped to a cell line or model system.
`draft_claim`, `author_claim`	Propose a claim (validate only) or commit one to your tenant's graph. Promotions are recorded as requests, never auto-approved.
`provenance`	Trace a derived result back to the recordings it came from.

Bring data in

Tool	What it does
`upload_recording_init`	Start a recording upload — returns presigned S3 URLs; bytes go straight to S3, never through the connector.

Other MCP clients

The same server works with any MCP client — for example Claude Desktop or Claude Code — by adding https://app.openculture.ai/mcp as a custom / remote MCP server and signing in with your Open Culture account.

Prefer to call the API yourself?

If you're building your own agent rather than using these tools, drive the Habitat REST API directly — see the From code examples.