Your workspace, inside your AI client.

The Levain MCP server puts your workspace (agents, runs, and analytics) one message away. Connect Claude, Cursor, or any Model Context Protocol client and start working in plain language.

Add to Claude Connect another client
Works with Claude Cursor Zed Continue VS Code Your own scripts

Talk to your workspace.

Ask in plain language. Your client picks the right tool and calls it as you. No dashboards, no context switching.

What did we spend on runs yesterday, by agent?

Using Levain MCP… get_billing get_run_costs

Yesterday's runs came to $48.20. Invoice Reconciler led at $22.10, then Support Triage ($14.60) and Lead Qualifier ($7.90). Want the hourly split?

Everyday asks.

You describe the outcome; your client reaches for the tools to get there. A few things teams run without opening the dashboard:

"Pause every schedule that failed overnight."

Enumerates your agents, lists each one's schedules, and pauses the ones that failed.

  • list_agents
  • list_schedules
  • pause_schedule

"How much did we spend last week, and which models drove it?"

Pulls your usage summary for the period, broken down by model and provider.

  • get_usage_summary

"Make a copy of Support Triage for the EMEA team."

Reads the agent's name and description, then creates a new one.

  • get_agent
  • create_agent

"Add our new refund policy to the knowledge base."

Finds the right page and writes the update into your wiki.

  • list_wiki_pages
  • update_wiki_page

51 tools, your whole workspace.

Everything you can do in Levain, grouped the way you reach for it. You never type these names; your client picks the right tool and shows you each write before it runs. The full list is here so you can see how much is on offer.

Read · write

Agents & versions

  • list_agents
  • get_agent
  • create_agent
  • update_agent
  • delete_agent
  • list_agent_versions
  • get_agent_version
  • list_templates
  • list_managed_agents
  • adopt_managed_agent
Read · write

Runs & sessions

  • create_run
  • list_runs
  • get_run
  • get_run_logs
  • get_run_log_events
  • get_run_costs
  • get_run_requests
  • list_sessions
  • get_session
  • get_session_events
  • get_session_state
  • cancel_session
Read · write

Schedules & triggers

  • create_schedule
  • list_schedules
  • get_schedule
  • pause_schedule
  • resume_schedule
  • delete_schedule
  • create_agent_trigger
  • list_agent_triggers
  • delete_agent_trigger
Read · write

Knowledge base

  • get_current_wiki
  • list_wiki_pages
  • get_wiki_page
  • list_wiki_sources
  • update_wiki_page
  • delete_wiki_page
Read

Analytics & billing

  • get_usage_summary
  • get_billing
  • get_hourly_usage
  • list_activity
Read

Integrations

  • list_integrations
  • get_integration
  • list_agent_integrations
  • list_integration_providers
Read · write

Notifications

  • list_notifications
  • mark_notification_read
  • mark_all_notifications_read
Read

Workspace & sandboxes

  • whoami
  • list_sandboxes
  • get_sandbox

Resources

Stable documents your client can pull in directly, without spending conversation context: recipe specs at recipe://<slug> and the full OpenAPI reference at schema://openapi.

Prompts

Built-in workflows your client exposes as slash-commands, for the things you do often: getting started with your first agent, or debugging a stalled run.

Yours, and only yours.

The server acts as you, scoped to your workspace, on the same audit trail as everything else on the platform.

Sign in once

Your client opens a browser once. You sign in, tokens are issued for your workspace, and there's nothing to paste, store, or rotate. Revoke access any time.

Your role still applies

Every call runs as the signed-in user. The server only ever sees your workspace, and it respects your role, so a member can't reach what an admin can.

Nothing happens off the record

Reads and writes are metered and audit-logged like anything else on Levain. What your client did, and when, is always there to inspect.

Common questions

What is MCP?
Model Context Protocol is an open standard that lets AI clients discover and call tools on your behalf. Levain exposes your workspace over MCP, so clients like Claude and Cursor can work with your agents, runs, and analytics directly, with no custom integration to build.
Which clients does it work with?
Any client that speaks Model Context Protocol: Claude Desktop and web, Cursor, Zed, Continue, VS Code, or your own scripts. The connection is a standard HTTP MCP endpoint, so anything built to the spec works.
Do I need an API key?
No. Authentication is a standard OAuth browser flow. Your client opens a sign-in once, tokens are issued for your workspace, and there's nothing to copy or manage.
Can it change things, or is it read-only?
Both. Read tools pull data: runs, costs, logs, the knowledge base. Write tools create agents, start runs, manage schedules, and add provider keys. Your client shows you each write before it runs, so nothing changes without your say-so.
How is access scoped and secured?
Every call runs as the signed-in user, scoped to that user's workspace and role. The same permissions that govern the app govern the MCP server, and every action lands on the platform audit trail.
What does it cost to use?
The MCP server itself is part of the platform. You're billed the same way as always: your plan plus usage for the LLM volume your agents run. Reading analytics or listing agents doesn't cost anything; starting a run costs what that run costs. See pricing for the detail.
How is this different from connecting an MCP server to my agents?
Opposite directions. This page is about pointing your own AI client at Levain to manage the platform. Giving your agents new tools (connecting Notion, Slack, or a custom MCP server so an agent can use them mid-run) is integrations.

Connect in a minute.

One URL, a browser sign-in, and your workspace shows up as tools your client can call. No API keys to copy or rotate.

Claude

One click. Claude opens with the Levain connector prefilled, so you sign in once and your workspace is in every chat, on desktop or web.

Add to Claude OAuth · nothing to paste

Cursor, Zed & any client

Add Levain as a custom MCP server, paste the URL, and sign in through your browser. The Levain tools then show up in chat.

Server https://mcp.levainlabs.com
Cursor Zed Continue VS Code ChatGPT Your scripts
ChatGPT needs Developer Mode

Your workspace belongs in your AI client.

One click adds Levain to Claude. On any other client, paste the server URL and sign in; your workspace is a message away.

Add to Claude Read the docs