MCP
Your coding agent already reads your code. Let it read your users.
One command, an OAuth sign-in, and your editor can ask what actually happened.
Read the MCP docs
Clado
Olive
Freebuff
Attensira
Assemble
Opennote
Stratify
Gatekeep- Clado
- Olive
- Freebuff
- Attensira
- Assemble
- Opennote
- Stratify
- Gatekeep
- Clado
- Olive
- Freebuff
- Attensira
- Assemble
- Opennote
- Stratify
- Gatekeep
What it can reach
Thirteen tools your agent picks between itself.
You rarely name a tool. You ask a question in the editor and the agent chooses — but the permissions you ticked at connect time decide which tools exist for that connection at all.
Product analytics
Events, sessions, visitors, and the event schema, queried with SQL against your own project rather than through a filter picker.
Replays as text
A session arrives as a timeline of navigations, clicks, on-screen messages and detected friction — every line carrying a replay link seeked to that moment.
Issues
Crash, error and UX-friction issues with their stacks, breadcrumbs and affected users, so the fix starts from what broke rather than from a title.
Friction on one page
Rage clicks, dead clicks, stuck spinners and abandoned forms on a route, ranked by how many sessions hit each one.
Funnels
Compute a funnel over any step sequence, and — with the permission granted — save it so it appears on your dashboard.
Triage
Resolve, ignore, mute, reopen or assign an issue. Applied changes are attributed to an agent, not to a person clicking a button.
Connect
One command. No API key.
Your editor opens a browser, you sign in, pick a project, and tick what the agent may do. Nothing sensitive lands in a config file you might commit.
# Add the server
claude mcp add --transport http humanbehavior \
https://www.humanbehavior.co/api/mcp/mcp
# Sign in — your browser opens the consent screen
claude mcp login humanbehavior{
"mcpServers": {
"humanbehavior": {
"url": "https://www.humanbehavior.co/api/mcp/mcp"
}
}
}How it behaves
It cannot see more than you can.
- 01
Every call re-checks your membership
Connecting an editor does not widen your access: each request is verified against your own project membership, and a connection can be pinned to a single project.
- 02
Writes are previewed, then confirmed
Ask an agent to resolve an issue and it returns the change, not the result. You approve that specific change; retries with the same idempotency key apply once.
- 03
End-user content is labelled as data
Text your users typed arrives inside an untrusted block, so the agent reports directions found in it instead of following them.
Details
The parts worth knowing before you connect.
- Endpoint
- https://www.humanbehavior.co/api/mcp/mcp — streamable HTTP, request/response. There is no SSE stream. We test each release against Claude Code; Cursor and Codex speak the same standard OAuth flow and work.
- Permissions
- Read product analytics, read session replays as text, read issues, triage issues, and save funnels. Grant only the reads and every write tool disappears from the list the agent is shown.
- Never available
- Deleting projects, purging data, rotating API keys, managing members, and billing are out of reach of any connection at any permission level.
- Disconnecting
- Settings → MCP lists each connection, what it can do, and when it was last used. Revoking one takes effect immediately, including for an editor session that is open right now.
- Limits
- 600 reads a minute and 60 writes every five minutes, counted per person. Query results cap at 200 rows and every query carries a 30-second database ceiling, so a runaway loop fails fast instead of hanging your editor.
Keep going
The rest of the platform.
Everything here reads from the same recorded session — there is nothing to stitch together between surfaces.
Stop guessing what happened.
One SDK, one afternoon. We will walk your team through it on a call and leave you with the recording of your own product.