Skip to main content
Use implementation prompts when you want an AI coding agent to pick up a feedback request and implement it. Project Echo turns the request into agent-ready Markdown using Anthropic Claude Sonnet (default API model id: claude-sonnet-4-5). You can generate prompts in the staff app (see How to generate an implementation prompt), or automate with the REST API and MCP.

​
What the model uses

Claude receives:
  • Request title and description
  • Public comments
  • Internal notes (staff-only)
The stored prompt may therefore include internal-note content. Anyone with requests:read can later read that prompt via API/MCP.

​
Permissions

Create or update a key under API & MCP. Existing keys do not gain new abilities automatically—edit the key or create a new one (Contributor preset includes requests:update).

​
REST API

Base URL: https://api.projectecho.io/v1 Authenticate with both headers on every call:
  • Authorization: Bearer pe_api_…
  • X-PE-API-Key-Id: <key-uuid>

​
Get the stored prompt

Returns { "prompt", "generated_at", "generated_by" }, or 404 if none exists yet.

​
Generate or regenerate

Overwrites any previous prompt and returns the full Markdown. Soft rate limit: about 30 generations per workspace per hour.

​
Discover readiness on get_request

GET /v1/requests/:id includes:
  • implementation_prompt
  • implementation_prompt_generated_at
  • implementation_prompt_generated_by
List responses include implementation_prompt_generated_at (without the full prompt body).

​
MCP tools

With MCP connected (@project-echo/mcp 0.2.1+): Both take request_id. Example (Cursor / Claude agent):
If those tools are missing after an upgrade, reinstall @project-echo/mcp@latest and restart the MCP server in your client.

​
Model (Anthropic)

Echo calls Anthropic’s Messages API on the server. API keys for Anthropic are not configured in your MCP client—only your Project Echo API key is.

​
Privacy and safety notes

  • Generated prompts can embed internal notes. Treat stored prompts like staff-sensitive content.
  • Prefer least-privilege keys; only grant requests:update to agents that should regenerate prompts.
  • Generation is rate-limited to reduce accidental loops.