Early access

Clarabit Docs

Clarabit is a hosted MCP server plus a small web app. Connect it once to the AI tools you already use, and they all read and write the same tasks, pages, context, and agent sessions. There is nothing to run locally and no API key to keep in a dotfile.

Install

One connector URL works everywhere: https://mcp.clarabit.ai/mcp. Pick your host below. First create your workspace, then follow the setup checklist to connect your AI app. Until Clarabit is listed in the Claude and ChatGPT directories, it is added as a custom connector.

Claude

Works on claude.ai, Claude Desktop, and mobile.

1Open Claude, go to Settings → Connectors, and choose Add custom connector.
2Name it Clarabit and paste the connector URL below as the server URL.
3Click Connect and sign in with Google when Claude asks. Then say: "What's waiting on me in Clarabit?"

Connector URL

https://mcp.clarabit.ai/mcp

ChatGPT

Available on ChatGPT web; needs Developer mode until our directory listing is live.

1In ChatGPT, open Settings → Apps & Connectors and turn on Developer mode (under Advanced).
2Choose Create, name it Clarabit, and paste the connector URL below as the MCP server URL. Authentication: OAuth.
3Sign in with Google when asked. In a new chat, add the Clarabit connector and say: "What's waiting on me?"

Connector URL

https://mcp.clarabit.ai/mcp

Claude Code

Terminal, VS Code, or JetBrains. Same connector, no API key.

1Add the connector from your terminal:

Command

claude mcp add --transport http clarabit https://mcp.clarabit.ai/mcp

2In Claude Code, run /mcp, pick clarabit, and sign in with Google in the browser that opens.
3Say: "Bootstrap Clarabit and show me what's assigned to me."

Connector URL

https://mcp.clarabit.ai/mcp

Codex, Cursor, and other MCP hosts

Any host that supports remote MCP servers (Streamable HTTP) with OAuth can connect with the URL alone. The host registers itself with Clarabit automatically (dynamic client registration) and opens a Google sign-in the first time it connects. The snippets below are the standard config locations; if your host's version differs, use its MCP settings and paste the URL.

Codex

Add to ~/.codex/config.toml, then start Codex and complete the sign-in when prompted:

[mcp_servers.clarabit]
url = "https://mcp.clarabit.ai/mcp"

Cursor

Add to .cursor/mcp.json in your project (or ~/.cursor/mcp.json for every project), then approve the connection in Cursor's MCP settings:

{
  "mcpServers": {
    "clarabit": { "url": "https://mcp.clarabit.ai/mcp" }
  }
}

Local coding agents with an API key

The npm package @limeadelabs/clarabit-mcp runs the same tools over stdio for agents that cannot use OAuth, plus a few local-only tools (session files, file uploads). See the package README. The hosted connector is the recommended path for everyone else.

What OAuth grants

When you connect, your AI tool asks Clarabit for a token with two scopes:

Scope	What it allows
`clarabit:read`	Read your projects, tasks, pages and their versions, comments, context entries, labels, and the workflow. Everything in the Read group below.
`clarabit:write`	Create and update tasks, pages, comments, and context; claim and move tasks through the workflow; record agent sessions. Everything in the Write and Session groups below.

The token identifies you. The tool acts under your permissions and can only do what you could do in the web app; a task in a workspace you are not a member of is simply not there.
Every write is recorded under your identity, marked as made via an agent, so a person's edits and an agent's edits are always distinguishable.
Remove the connector from your AI app and revoke its connection in Clarabit to end access. Disconnecting Google sign-in is separate from revoking an AI app connection.
Clarabit never sees your conversation, only the tool calls the host makes. See the privacy policy for what is logged about each call.

Tools

All tools are prefixed cla_. Your AI tool will usually pick the right one; you rarely need to name them. Start any session with cla_bootstrap: it returns your projects, what is assigned to you, and what needs your attention, then ask it to read the instructions for the project you want to work in.

Read

Tool	What it does
`cla_bootstrap`	Orientation for a new session: your projects, your assignments, what is waiting on you.
`cla_list_projects` `cla_get_project`	Projects in your workspace, and one project's details and settings.
`cla_list_tasks` `cla_get_task`	Tasks by project or status; one task with its description, comments, links, and spec.
`cla_get_workflow`	The task workflow: which status transitions are allowed, and which need a claim, a comment, or a human.
`cla_generate_prompt`	A ready-to-use working prompt for a task, assembled from its spec and context.
`cla_list_pages` `cla_get_page`	Pages (specs, decisions, drafts) in a project, and one page's full content.
`cla_list_page_versions` `cla_get_page_version`	A page's version history, and any earlier version.
`cla_list_page_comments` `cla_prepare_page_comment_anchor`	Comments on a page, and the anchor needed to comment on an exact passage.
`cla_context_list` `cla_context_get`	Project context entries (the durable background every session should know).
`cla_context_package`	The assembled context package for one task: instructions, spec, pages, and comments, in one call.
`cla_list_labels`	Labels available in a project.

Write

Tool	What it does
`cla_create_task` `cla_update_task`	Create a task; change its status, priority, title, description, or assignee. Status changes are checked against the workflow.
`cla_start_task` `cla_submit_task`	Move a task to in progress; hand it to review with the evidence attached.
`cla_claim_task` `cla_release_task`	Claim a task so no other session builds the same thing; release it again.
`cla_add_comment`	Comment on a task: progress, questions, findings.
`cla_add_label_to_task` `cla_remove_label_from_task`	Label and unlabel tasks.
`cla_log_time`	Log time against a task, for a person or an agent.
`cla_create_page` `cla_update_page`	Create and revise pages. Content changes made through the API create a new version that you can review or restore.
`cla_archive_page` `cla_restore_page` `cla_restore_page_version`	Archive and restore a page, or roll it back to an earlier version.
`cla_create_page_comment` `cla_update_page_comment`	Comment on an exact passage of a page, and edit that comment.
`cla_resolve_page_comment` `cla_unresolve_page_comment`	Mark a page comment as addressed, or reopen it.
`cla_context_create` `cla_context_update`	Add or update a project context entry.
`cla_upload_image` `cla_upload_task_attachment` `cla_upload_comment_attachment`	Attach files and screenshots. Local coding agents only for now; hosted clients (Claude, ChatGPT) do not yet get file upload.

Session

Sessions are the durable record of an agent run: which agent, which task, what happened. Agents open one when they pick a task up and close it when they stop.

Tool	What it does
`cla_session_start`	Open a recorded run against a task.
`cla_session_progress`	Log a milestone as the work happens. This is what the activity log in the app shows.
`cla_session_event`	Log a structured event: a commit, a CI result, a merge request.
`cla_session_blocked`	Report a blocker so it shows up in someone's attention queue.
`cla_session_heartbeat` `cla_heartbeat`	Liveness pings for a session and for a claim, so a stalled run can be spotted.
`cla_session_complete` `cla_session_fail`	Close the run with a summary, as done or as failed.

Review and approval

Ask your AI app to save work for your review. Project workflows define which transitions need approval; cla_get_workflow explains the current rules. A connection acts with the permissions you grant it. Review important work before publishing or using it.

Pick up where you left off

Ask your AI app to save the task, its result, and the next step in Clarabit. In your next conversation, ask it to open that task and continue. It can read the saved record without you copying the whole conversation. Saving a task does not start a background worker: work happens when you ask your AI app.

To reduce reminders, save the project instructions suggested in Clarabit's setup checklist in your AI app's project settings. A prompt in one conversation does not configure every future chat.

Support and changes

Email support@clarabit.ai. A person reads it. Tell us which host you are using and what you asked it to do; a screenshot of the tool call helps.

For connector changes and upgrade instructions, read the MCP changelog. Product changes are in the public changelog. For how we handle your data, see the privacy policy and terms.