> ## Documentation Index
> Fetch the complete documentation index at: https://docs.observerbee.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Observerbee in your AI assistant

> Use Observerbee from Claude, ChatGPT, Cursor, Claude Code and Codex through the Observerbee MCP server.

Observerbee ships a remote [Model Context Protocol](https://modelcontextprotocol.io) server. Any MCP client can read your friction findings, journeys, screens, session evidence and task specs. Clients that support the MCP Apps extension (Claude, ChatGPT, Cursor, VS Code Copilot, Goose) also render interactive views inside the conversation: a findings board, an insight card with next steps, a journey rail, a screen inspector, a task spec review and the approval queue.

<CardGroup cols={2}>
  <Card title="Claude and ChatGPT" icon="message" href="/integrations/mcp-chat">
    Connect with OAuth. Ask what is hurting checkout, open a finding, dispatch it to your engineer agent and approve the pull request without leaving the chat.
  </Card>

  <Card title="Claude Code, Codex and Cursor" icon="terminal" href="/integrations/mcp-coding-agents">
    Connect with an MCP token. Pull an agent-ready task spec, implement it, and verify the fix against production data.
  </Card>
</CardGroup>

## What the server exposes

| Tool                                         | What it does                                                   | Interactive view |
| -------------------------------------------- | -------------------------------------------------------------- | ---------------- |
| `list_friction_findings`                     | Ranked friction findings for a project                         | Findings board   |
| `get_insight`                                | One finding with evidence, task specs, requests and next steps | Insight card     |
| `list_journeys`, `get_journey`               | Mined user journeys and the step rail                          | Journey rail     |
| `get_state_context`                          | One screen or loop: screenshot, traffic, transitions           | Screen inspector |
| `get_session_evidence`                       | Session metadata and the replay link                           |                  |
| `list_task_specs`, `get_task_spec`           | Agent-ready task specs as Markdown                             | Task spec review |
| `create_task_spec`                           | Compose a spec from a finding (idempotent)                     |                  |
| `list_action_requests`, `get_action_request` | Governed proposals waiting for a decision                      | Approval queue   |
| `verify_fix`                                 | Compare production metrics with the spec's baseline            |                  |

Human decisions (`review_action_request`, `set_task_spec_status`, `promote_insight_to_triage`, `request_implementation`, `set_insight_status`) are only callable from the views. The model can describe them but never trigger them; a person clicks.

Every decision made through an assistant is attributed to the signed-in member and written to the same accountability trail as the dashboard.

## Resources and prompts

* `observerbee://task-specs/{id}` and `observerbee://insights/{id}`: Markdown documents for clients that attach context instead of calling tools.
* `fix_friction`: a guarded implementation brief for a coding agent, ending with a `verify_fix` step.
* `triage_findings`: ranks a project's open findings and proposes which to turn into specs.