Latch MCP

Latch provides a remote Model Context Protocol server:

https://mcp.latch.bio/mcp

It lets compatible AI clients discover workflows and interact with Latch Data and executions through an OAuth-authorized tool surface.

Authentication Boundary

• A Latch account is required.
• MCP uses OAuth in the connected AI client.
• MCP authorization does not authenticate the Latch SDK, CLI, or Console.
• SDK login credentials do not replace MCP authorization.
• Actions launched through MCP incur the same normal Latch charges as the corresponding Console or SDK action.

Never copy OAuth tokens between these authentication domains.

Setup

Cursor

Add:

{
  "mcpServers": {
    "LatchBio": {
      "url": "https://mcp.latch.bio/mcp"
    }
  }
}

Then open Cursor Settings → Tools & MCP(s), select LatchBio, click Connect, and complete OAuth.

Claude Code

claude mcp add --transport http latchbio https://mcp.latch.bio/mcp

Run /mcp and complete OAuth.

Codex

codex mcp add latchbio --url https://mcp.latch.bio/mcp
codex mcp login latchbio

For another MCP client, configure the same remote HTTP URL and follow that client's OAuth flow.

Documented Tools

The current guide lists:

Tool names and schemas can evolve. Discover the connected server's current tools before calling them.

Safe Launch Workflow

1. Select a workspace
2. Call list_workspaces.
3. Prefer the default only when it is clearly the intended target.

1. Discover rather than guess
2. Call list_workflows.
3. Identify the exact workflow and version.

1. Fetch the schema
2. Call get_workflow_schema.
3. Use returned types, required values, enums, defaults, and path rules.

1. Resolve data
2. Use list_files only for the minimum directories needed.
3. It lists metadata, not file contents.
4. Use get_file for a selected file's access link. Prefer its Console link for sharing; treat a presigned download URL as a temporary credential.
5. Do not browse unrelated or sensitive directories.

1. Prepare a launch summary
2. Workspace
3. Workflow and version
4. Input paths and parameters
5. Output destination
6. Expected resource/cost implications

1. Confirm
2. Obtain user approval before launching paid compute.
3. Reconfirm unusually large, GPU, or fan-out workloads.

1. Launch
2. Call launch_workflow once.
3. Preserve the returned execution identifier and Console URL.

1. Monitor
2. Poll with get_execution.
3. Use get_task_logs only for relevant nodes.
4. Avoid repeatedly downloading full logs when an inline excerpt is enough.

Example Agent Plan

list_workspaces
→ choose workspace 12345
→ list_workflows
→ choose Bulk RNA-seq version X
→ get_workflow_schema
→ validate sample paths and output directory
→ present launch summary and request confirmation
→ launch_workflow
→ get_execution until terminal status
→ get_task_logs only if a task fails

Cost and Data Safety

• Listing resources is read-only; launching is not.
• Do not launch during exploratory discovery.
• Do not expose signed log URLs; they may grant temporary access.
• Do not paste secrets into workflow parameters.
• Confirm paths belong to the selected workspace.
• Avoid broad file listing when exact paths are already known.
• Stop polling after a terminal state.
• Scientific validation still requires inspecting outputs and method assumptions after orchestration succeeds.

When MCP Is Unavailable

Use:

• Latch Console for interactive discovery and launch
• latch_cli.services.launch.launch_v2 for Python automation
• latch ls or LPath for data inspection
• Latch Console for execution monitoring; latch get-executions remains in 2.77.1 but is deprecated and scheduled for removal

Do not simulate a missing MCP tool by inventing undocumented HTTP endpoints.

Official Source

• Latch MCP guide: https://wiki.latch.bio/agent/latch-mcp