> ## 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.

# Claude Code, Codex and Cursor

> Give your coding agent Observerbee's evidence and task specs, and verify fixes against production.

## Connect

Mint an MCP token under Settings, API Keys in the dashboard. Tokens are secret; treat them like passwords. Org-wide tokens read every project; project-scoped tokens read one.

<Tabs>
  <Tab title="Claude Code">
    ```bash theme={null}
    claude mcp add --transport http observerbee https://mcp.observerbee.com/mcp --header "Authorization: Bearer YOUR_MCP_TOKEN"
    ```
  </Tab>

  <Tab title="Codex">
    ```bash theme={null}
    export OBSERVERBEE_MCP_TOKEN=YOUR_MCP_TOKEN
    codex mcp add observerbee --url https://mcp.observerbee.com/mcp --bearer-token-env-var OBSERVERBEE_MCP_TOKEN
    ```
  </Tab>

  <Tab title="Cursor">
    Add to `.cursor/mcp.json` (project) or `~/.cursor/mcp.json` (global):

    ```json theme={null}
    {
      "mcpServers": {
        "observerbee": {
          "url": "https://mcp.observerbee.com/mcp",
          "headers": { "Authorization": "Bearer YOUR_MCP_TOKEN" }
        }
      }
    }
    ```
  </Tab>
</Tabs>

## The loop

<Steps>
  <Step title="Pick a spec">
    Ask the agent to list task specs, or attach `observerbee://task-specs/{id}` as context. Specs are composed from findings by a person in the dashboard or in a chat host and marked ready when reviewed.
  </Step>

  <Step title="Implement with the guarded prompt">
    Use the `fix_friction` prompt. It tells the agent to confirm the defect in the source before changing anything, to open a draft pull request only, to cite the evidence links, and to treat each acceptance criterion as a checklist item.
  </Step>

  <Step title="Verify in production">
    After the change is deployed and has collected traffic, ask the agent to run `verify_fix` with the task spec id. It compares outbound transitions from the affected screen, drop-off at the affected journey step, journey health and loop traffic with the baseline captured when the spec was composed, and reports a conservative verdict per check.
  </Step>
</Steps>

Verdicts are measurements, not gates: "unchanged" often means the fix has not collected enough sessions yet. Run it again later.

## Tools your agent will use

`list_task_specs`, `get_task_spec`, `get_insight`, `get_journey`, `get_state_context`, `get_session_evidence` and `verify_fix`. All are read-only. Decisions (marking a spec ready, approving a request) stay with people in the dashboard or a chat host.