MCP server

The Skynet MCP server lets an AI agent use Skynet for you. Claude, ChatGPT, Codex, Cursor and other MCP clients can check a run setup, start runs, watch them and stop them. You approve the connection one time in your browser. The agent sets a spending limit for each run.

Server URL

Every client uses the same URL:

https://mcp.skynetml.com/mcp

The server uses Streamable HTTP. Sign-in uses OAuth 2.1. You do not copy a token or an API key into the client.

Connect your client

Select your client. Follow its steps, then go to Approve the connection.

Last checked against each client's docs: 2026-11-01.

claude mcp add --transport http skynet https://mcp.skynetml.com/mcp

claude mcp add --transport http --scope user skynet https://mcp.skynetml.com/mcp

codex mcp add skynet --url https://mcp.skynetml.com/mcp
codex mcp login skynet

[mcp_servers.skynet]
url = "https://mcp.skynetml.com/mcp"

{
  "mcpServers": {
    "skynet": {
      "url": "https://mcp.skynetml.com/mcp"
    }
  }
}

code --add-mcp '{"name":"skynet","type":"http","url":"https://mcp.skynetml.com/mcp"}'

{
  "servers": {
    "skynet": {
      "type": "http",
      "url": "https://mcp.skynetml.com/mcp"
    }
  }
}

gemini mcp add --transport http skynet https://mcp.skynetml.com/mcp

{
  "mcpServers": {
    "skynet": {
      "httpUrl": "https://mcp.skynetml.com/mcp"
    }
  }
}

{
  "mcpServers": {
    "skynet": {
      "serverUrl": "https://mcp.skynetml.com/mcp"
    }
  }
}

https://mcp.skynetml.com/mcp

Approve the connection

The first time the agent calls Skynet, your browser opens a Skynet page.

1. Sign in to Skynet, if you are not signed in.
2. Read the page. It shows the agent name, the site it returns to, and your account.
3. Read what the agent can do:
   • See your optimization runs and their results.
   • Start and stop optimization runs, paid from your balance.
4. Click Allow to connect. Click Deny to stop.

After you click Allow, the browser returns to the agent. The agent can now call the tools below.

Tools

Read access uses the runs:read scope. Write access uses the runs:write scope.

No tool adds funds, changes your balance or changes billing settings.

How an agent runs an optimization

1. The agent calls check_run with the run setup. Skynet runs the static checks and returns:
   • The result of each check.
   • The lowest spending limit the run can start with.
   • A suggested spending limit. It is at least $5.00.
   • The models that have no known price.
   • The amount you can spend now.
2. The agent calls start_run with a spending limit (max_cost_cents, from 1 to 1,000,000 US cents). It can also send an idempotency_key, so that a retry does not start a second run.
3. If the setup check is still running, start_run returns in_progress and retry_after_seconds. The agent calls start_run again with the same input after that time.
4. The agent calls get_run until the run ends. The final statuses are success, failed, cancelled and stopped.

The run stops when it reaches its spending limit. Read Billing for how Skynet charges a run.

Security

• Each client gets its own grant. The client never sees your password.
• Sign-in uses PKCE (S256). The client registers itself with dynamic client registration or a client ID metadata document.
• Access tokens expire after 1 hour. The client renews them with a refresh token.
• A refresh token expires after 30 days without use. A grant lasts at most 90 days. Then you approve the connection again.
• Each refresh token works only one time. If someone uses an old refresh token again, Skynet ends the grant.
• The agent can only see and change your own runs.

Read Security for how Skynet protects your account and data.

Disconnect

To disconnect an agent, remove the Skynet server from the client. Most clients then revoke their tokens. If a client does not revoke them, its refresh token expires after 30 days without use.

Troubleshooting

For direct HTTP access without an agent, read API.