Audience: manage AgentStack projects from ChatGPT or Gemini chat windows.
Endpoint: https://agentstack.tech/mcp
Setup hub: MCP_SETUP_QUICKSTART.md
Terminal or any HTTP client:
curl -s -X POST https://agentstack.tech/mcp/tools/projects.create_project_anonymous \
-H "Content-Type: application/json" \
-d '{"params": {"name": "My AI Project"}}'The response returns user_api_key or project_api_key at the top level, plus neutral bootstrap metadata (which headers to set — no duplicated secrets in instructions).
Configure once in ChatGPT/Gemini settings or your password manager — paste the key into the connector’s API Key / Token field (X-API-Key header). Do not rely on the model to store it.
Signed-in users: create a scoped key at agentstack.tech/user/profile?tab=api instead.
Works with ChatGPT Plus and above.
- Open chatgpt.com → Explore GPTs → Create.
- Configure → Actions → import OpenAPI from the repo:
https://github.com/agentstacktech/gpt-plugin/blob/main/openapi/agentstack-mcp.yaml - Authentication → API Key:
- Header:
X-API-Key - Value: your key from step 0
- Header:
- Instructions — copy from GPT_INSTRUCTIONS.md.
- Save → “Only me” or “Anyone with link”.
Open your GPT and use natural-language commands (see cookbook below).
Requires ChatGPT Plus / Pro / Business with Developer Mode.
- chatgpt.com → Settings → Apps & Connectors → Advanced → enable Developer mode.
- Create connector:
- Name: AgentStack
- URL:
https://agentstack.tech/mcp - Authentication: Token → paste your API key (AgentStack also accepts OAuth)
- You should see 1 tool:
agentstack.execute.
In chat: + next to the input → More → select AgentStack.
Template JSON: chatgpt-mcp-connector.template.json
Google limitation: custom MCP in web Gemini requires Gemini Spark (US, 18+, personal Google account, Keep Activity on). Standard Gemini chat does not support custom MCP.
- gemini.google.com → Settings & help → Connected Apps.
- Enable Keep Activity under Settings → Activity.
- Custom apps for Spark → Add a custom app.
- MCP server URL:
https://agentstack.tech/mcp - Complete OAuth if prompted, or use API key credentials from step 0.
- Next → confirm.
In Spark chat: @ → your custom app → command.
No Spark? Use ChatGPT (option A) or Gemini CLI in the terminal.
| Goal | Example prompt |
|---|---|
| List projects | “Call AgentStack discovery.list and show my projects.” |
| Publish a page | “Use hosting.publish_static on project_id X with this HTML…” |
| Check catalog | “GET /mcp/actions/summary and tell me how many actions exist.” |
| Read-only bootstrap | “Run recipe mcp_read_bootstrap with continueOnError.” |
Full cookbook: MCP_CHAT_COMMAND_COOKBOOK.md
| Problem | Fix |
|---|---|
| 401 / unauthorized | Re-check API key in GPT/Gemini connector settings |
| Gemini missing custom app | Requires Gemini Spark (US); otherwise use ChatGPT |
| ChatGPT “Tool scan failed” | Retry save 2–3×; URL must be exactly https://agentstack.tech/mcp |
| Model invents data | Say: “Call AgentStack API — do not fabricate project IDs.” |
| Permission denied | Key has narrow service_caps — create a wider key at Profile → API keys |
Live probe (no local Core clone):
curl -sS -H "X-API-Key: $AGENTSTACK_API_KEY" https://agentstack.tech/mcp/actions | headOr use your IDE plugin Connect + /agentstack-diagnose.
| ChatGPT Custom GPT | ChatGPT MCP Connector | Gemini Spark | |
|---|---|---|---|
| Difficulty | ★★☆ | ★★★ | ★★★ |
| Plus / Spark required | Plus | Plus | Spark |
| Works in normal browser chat | ✅ | ✅ | Spark only |
| One-time setup | ✅ | ✅ | ✅ |
Recommendation: start with ChatGPT Custom GPT (option A) for the fastest path.
RU guide: MCP_BROWSER_QUICKSTART_RU.md · Matrix: MCP_CLIENT_COMPATIBILITY_MATRIX.md