Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

MCP Server

Connecting AI tools to Lab Atlas using model context protocol (MPC)

Lab Atlas exposes a Model Context Protocol (MCP) server so AI tools — Claude Code, claude.ai, and any other AI agent — can browse your projects, studies, assays, notes, and tasks the same way a human collaborator would. The MCP server reuses your existing Lab Atlas API key for authentication, so there's no separate signup, no OAuth dance, and no second password to manage.

This guide walks through getting a key, connecting an AI tool, and using the tools the MCP server exposes.

Lab Atlas MCP is currently available only to Enterprise customers.

The MCP server is currently in beta. Endpoints and entities may change over time.

What MCP gives you

Once you connect an AI tool to the Lab Atlas MCP server it can, on your behalf:

  • Find projects, studies, and assays you have access to.

  • Pull the latest content of any electronic-lab-notebook note — whether attached to a project, study, assay, or the organization itself.

  • Read the task list on a study or assay, including status and due dates.

  • Read the external links, collaborators, teams, keywords, and study collections in your organization.

  • See which integrations (S3, SharePoint, Benchling, GitLab, etc.) and storage drives are connected.

It does not currently create, modify, or delete anything in this release. An agent can summarize, search, and cross-reference, but it cannot change your data.

Prerequisites

You need an API key. The MCP server uses the exact same key format as the Lab Atlas REST API, so if you already have one for scripting or integrations you can reuse it.

To create one:

  1. Sign in to your Lab Atlas tenant.

  2. Open the user menu in the top-right corner and choose Account settings.

  3. Select the API keys tab.

  4. Click Create new key, give it a memorable name (e.g. claude-code-laptop), and pick an expiration.

  5. Copy the value from the modal. It looks like lak_<keyId>.<secret>. This is the only time the full value is shown — store it somewhere safe.

The key inherits your user's permissions and is scoped to your organization. An agent connecting with your key can see exactly what you can see in the web UI, no more and no less.

If you do not see the API keys tab, the feature is not enabled for your tenant — contact your administrator.

You also need the MCP server URL for your tenant. The path is always /mcp (eg. https://<your-tenant>.labatlas.com/mcp ).

Connect from Claude Code

Claude Code is Anthropic's terminal-based AI assistant. It manages MCP servers via the claude mcp subcommand.

The --transport http flag selects the Streamable HTTP transport, which is what the Lab Atlas server speaks. The --header is forwarded with every request.

After this, start a Claude Code session as you normally would (claude) and the labatlas tools will be available. Try:

"Use the labatlas mcp to list my projects."

To remove or refresh the registration:

Connect from claude.ai

claude.ai (the web app) supports remote MCP servers via the Connectors panel.

  1. Open Settings → Connectors.

  2. Choose Add custom connector (the exact label may say "Add custom MCP server").

  3. Fill in:

    • URL: https://your-tenant.labatlas.com/mcp

    • Transport: HTTP / Streamable HTTP (whichever option the UI offers)

    • Custom header: name X-API-KEY, value lak_xxxxxxxx.xxxxxxxxxxxxxxxxxxxx

  4. Save. The connector should report Connected.

Once connected, the Lab Atlas tools appear in the tool picker for any conversation, including Projects.

Connect from the Anthropic API

If you are building your own agent against the Anthropic API, configure the MCP server in the mcp_servers array on the message request:

The Anthropic API forwards authorization_token as Authorization: Bearer <token>. Lab Atlas does not accept bearer tokens for API key auth, so if the SDK you use does not let you set a custom header, use the explicit Authorization: ApiKey <key> form via a header field instead. Most SDKs support a headers map on the MCP server config:

Available tools

The server registers 52 read-only tools today, grouped by resource:

Resource
Tools

Projects

list_projects, get_project

Studies

list_studies, get_study, list_studies_by_project

Assays

list_assays, get_assay, list_assays_by_study

Assay types

list_assay_types, get_assay_type

Notes

list_organization_notes, list_project_notes, list_study_notes, list_assay_notes, get_organization_note, get_project_note, get_study_note, get_assay_note

Tasks

list_study_tasks, list_assay_tasks, get_task

External links

list_project_external_links, list_study_external_links, list_assay_external_links

Collaborators

list_collaborators, get_collaborator

Teams

list_teams, get_team

Keywords

list_keywords, get_keyword

Study collections

list_study_collections, get_study_collection

Study relationships

list_study_relationships

Integrations

list_integrations, get_integration

Storage drives

list_storage_drives, get_storage_drive

Storage folders

list_project_storage_folders, list_study_storage_folders, list_assay_storage_folders

Shared storage folders

list_shared_storage_folders, get_shared_storage_folder

Organization users

list_organization_users, get_organization_user

Users

list_users, get_user

Activity

list_activities, get_activity, list_activities_by_project, list_activities_by_study, list_activities_by_assay, list_activities_by_user

Listing tools accept optional page (zero-indexed) and size (max 100) parameters. Look-up tools take a single UUID parameter. The response shape mirrors the JSON returned by the equivalent GET /api/v1/... endpoint — see Public-API-Reference.md for field-level docs.

You normally do not pick tools by hand. Ask the agent for what you want and let it choose:

"Find every study tagged with the metabolomics keyword and summarize the most recent note on each."

The agent will plan a call sequence like list_keywords → list_studies → list_study_notes → get_study_note and assemble the answer.

Example prompts

A few prompts that exercise the read surface:

  • "List my active projects and tell me which ones have studies that were updated this week."

  • "In project CPA, summarize the open tasks across all studies and group by assignee."

  • "Pull the latest note from study TID-101 and extract the experimental conditions table."

  • "For every assay in study PPB-101, list its external links and check whether any point to deprecated tools."

  • "Show me which integrations are configured and what storage drives back the project CPA-100."

Privacy and scope

  • Authentication. Every request authenticates via your API key. The MCP server has no anonymous mode.

  • Authorization. The agent inherits your permissions. It cannot see anything you cannot see in the web UI. Cross-organization reads are not possible — the server rejects them with a 404-equivalent error.

  • Audit. Tool calls run as your user and are visible in the same activity feeds and integration audit logs that other API traffic produces. See integration-audit-logging-guide.md for details.

  • Data flow. Tool responses are JSON. They leave Lab Atlas only over the same HTTPS channel that serves the REST API, and only as part of a response to a request your agent made.

  • Revocation. Revoking the API key in the Account settings → API keys tab takes effect immediately and stops the agent in mid-conversation if a request is in flight.

If you are giving an agent access for a one-off task — say, a Claude Code session on a borrowed laptop — issue a short-lived key (one to seven days) instead of reusing your long-lived one.

Troubleshooting

SDK auth failed: HTTP 401: Invalid OAuth error response You're connecting with the older SSE transport, which Claude Code's SDK does not forward custom headers to. Use --transport http (the Streamable HTTP variant) as shown above.

401 Unauthorized on a tool call The key is missing, malformed, revoked, or expired. Generate a fresh key and reconnect.

Tool returns "X not found" The record is in another organization, or your user doesn't have view permission. The agent treats this as a normal "not found" — it is not a server error.

Tool list is empty The MCP server is not enabled in your deployment. Ask your administrator to check the spring.ai.mcp.server.enabled flag (see MCP-Configuration.md).

Connection seems slow on large pages Tool list responses are paginated to a max of 100 items. Ask the agent to scope its queries (e.g. by project) rather than scanning the full org.

For deeper issues, capture the agent's stderr or the Lab Atlas server logs and share them with your administrator.

PreviousREST API Reference

Last updated