Skip to main content

​
Overview

Hobbes runs a remote Model Context Protocol server at:
Connect an MCP-compatible AI client, such as Claude or Claude Code, and it can query Hobbes data, create Custom Links, update the Playbook in staging, and publish an approved agent version. The tools available to the client depend on the permissions you approve. Access is authorized with OAuth, not API keys. You sign in with your Hobbes account, choose one organization, and approve a specific set of permissions. The client only ever sees what you approved.
MCP connections are personal. Each teammate connects their own client and approves their own access. To integrate a backend system instead, use an API key.

​
Connect from Claude

1

Add the connector

In Claude, open Settings → Connectors → Add custom connector and enter https://api.hihobbes.com/mcp.
2

Sign in to Hobbes

Claude opens the Hobbes sign-in page in your browser. Use your normal Hobbes account.
3

Approve access

Pick the organization this connection may access, review the permissions, and click Approve. Only an organization admin can approve Custom Link creation, Playbook access, Playbook updates, or agent publishing.
4

Start asking

Back in Claude, the Hobbes tools are available immediately. Try: “Which of my demo sessions this week qualified, and why?”
On a Claude Team or Enterprise plan, custom connectors are managed by a workspace admin — ask them to add the connector for your workspace.

​
Connect from Claude Code

Register the server, then authenticate inside a session:
Run /mcp in Claude Code, select hobbes, and choose Authenticate. The same browser sign-in and consent flow applies. Once connected, Claude Code can pull Hobbes data into any task — reports, CRM scripts, prospect research.

​
Connect from other clients

Any client that supports remote MCP servers over streamable HTTP with OAuth can connect using the same URL — in Cursor, Windsurf, or VS Code, add it wherever the client configures MCP servers. For clients that can’t connect to remote MCP servers directly — they only launch MCP servers as local processes (stdio), or their remote support lacks OAuth (Codex, at the time of writing) — use the standard mcp-remote bridge. It runs locally as a small proxy and forwards everything to the Hobbes server:
The bridge opens the browser sign-in on first use and handles tokens from then on.
Connections made through the bridge appear as MCP CLI Proxy on the consent screen and in your connected clients list — that is the bridge’s own registered name, not the client behind it.

​
Permissions and tools

Permissions are chosen on the consent screen and fixed for the life of the connection. The client’s tool list shows only the tools its permissions allow. If you connected before Playbook and publishing permissions were available, revoke the old connection and reconnect. Select each new permission you want to grant. Existing connections never receive new permissions automatically. Every tool is scoped to the organization approved at connection time. The organization is derived from the connection, never accepted from the client. Some clients add their own approval prompt before running any tool — Claude Code’s “Do you want to proceed?” or Claude’s tool permission dialog. Those prompts belong to the client, not Hobbes; the permissions you approved at consent apply either way.
Transcripts contain the verbatim words of everyone in a session. Grant transcripts:read deliberately, and see Transcripts and security before piping transcript text into other systems.
custom_links:write allows a client to create personalized links:
  • Only an organization admin can approve it on the consent screen.
  • Batches are capped at 200 links per call and run as asynchronous jobs — the client receives a job it can poll with get_custom_link_job.
  • Requests are idempotent and quota-checked, so a retried or repeated call cannot double-create links. Large batches can take a couple of minutes to be accepted — a client that times out and retries with the same operation_key receives the original job instead of duplicates.

​
Update the Playbook

Playbook updates follow a preview-first process:
  1. The client reads the current staging and production Playbooks with get_playbook.
  2. preview_playbook_update validates the proposed changes and returns a before-and-after summary without changing staging.
  3. After you review the summary, apply_playbook_update applies that exact preview to staging. Previews expire after 24 hours, and a change to staging invalidates an older preview.
  4. Open the staging link returned by Hobbes and test the updated agent.
The client can also list ready Hobbes suggestions. Applying a suggestion uses the exact stored change. Use the preview flow when you want to edit the wording first.
Legacy Playbooks can be read through MCP, but they cannot be updated until they are migrated in Hobbes.

​
Publish an agent version

Publishing requires the separate agent:publish permission. Hobbes always:
  • asks for confirmation in the AI client before creating the publish job;
  • publishes the exact current staging version, not an inferred or older version;
  • runs the existing conversation tests before changing production; and
  • returns a job that the client checks until it succeeds or fails.
Grant agent:publish only when the connection should be able to change the production agent. Playbook updates can remain limited to staging by granting playbook:write without agent:publish. If you want a connection that cannot change data, leave all write and publishing permissions unselected.

​
Manage and revoke connections

Every client you approve appears in Hobbes → Account Settings → Connected AI clients, showing its name, organization, permissions, and connection date. Revoking a connection cuts the client’s access on its next request — before its token expires. Reconnecting later walks through sign-in and consent again.
Removing a connector inside the AI client does not revoke its access — the connection stays valid until you revoke it here. This page is the source of truth for what can reach your data.
Leaving an organization revokes access automatically: tools fail for that org’s connections the moment your membership ends, even mid-session.

​
Limits and errors

  • Requests are rate-limited per connection (HTTP 429 when exceeded); clients back off and retry automatically.
  • A 401 means the token expired or the connection was revoked — reconnect from the client.
  • A 403 means the connection lacks the permission for that tool, or your organization membership changed.
  • A connector that shows no tools available after being re-added was likely revoked in between — some clients reuse their stored credentials silently instead of prompting. Refresh the connector’s tool list or reconnect to re-authenticate.

​
MCP or API key?

Both surfaces expose the same data with the same organization scoping — pick per use case, or use both.