Skip to main content

AI agents for satellite operations

An AI agent connected to Constellation over the Model Context Protocol (MCP) can read your fleet, ask for forecasts and run workflows. It uses an API key like any other client, so it works inside the same scopes, plans and tenant isolation as the REST API, and you revoke it the same way.

What an agent can do​

ToolWhat it does
get_topologyRead the fleet as declared and as observed
get_predictionsForecasts at a time you name, from the models your deployment serves
get_telemetry, send_telemetryRead and write link records
get_network_config, patch_network_configRead or change the network config
search_catalogFind public satellites and ground stations to build the fleet from
list_workflows, put_workflow, create_scenario, run_workflow, get_runSave and run workflows (Automations in the Platform), optionally against a what-if scenario, and read their results
get_passesList the passes of your fleet between two times
list_actions, approve_action, reject_actionList recommendations held for approval and decide them; approving needs Operator + Act and the act:approve scope
load_demo_fleet, get_usageStart from a demo fleet, check this month's usage

The workflow tools appear only where workflows are on. The full list, with the arguments and scope each tool needs, is in Connect an AI agent.

What to ask​

  • "Load the demo fleet, then tell me which antennas will see each satellite ten minutes from now, and at what SNR."
  • "Forecast SNR for the links at ksat-svalbard five minutes from now and flag anything under 10 dB."
  • "Which of my ground stations are degraded right now?"

The limits it works inside​

  • Scopes. A key with only read scopes gives a read-only agent. Writing telemetry, the network config or workflows each needs its own scope.
  • Usage. Every tool runs a REST route with your key, so it is counted as the API is: a forecast call that returns at least one forecast counts as one prediction. No plan caps that count.
  • What-ifs stay what-ifs. A scenario run never delivers to a connector or destination, so an agent exploring options cannot reach an operator's screen with them.
  • Revocation. Revoke the key and the agent stops working everywhere.

Set it up​

The server is at https://api.constellation.space/mcp, over Streamable HTTP, with Authorization: Bearer <key>. Only /mcp takes the key that way; REST calls send it in x-api-key. Connect an AI agent has the configuration for each client.

Common questions​

Which AI clients can connect to the Constellation MCP server?

Claude Code, Claude Desktop, Cursor, VS Code, Codex CLI, Gemini CLI and Grok Build. Web connectors that sign in with OAuth, such as ChatGPT and claude.ai, are not supported yet.

Can an agent change my fleet?

Only within its key scopes. Give the agent a key with read scopes alone and it can read topology and forecasts but write nothing.

Does using an agent cost more than the API?

No. It is counted the same way: a get_predictions call that returns at least one forecast counts as one prediction, telemetry writes count as telemetry points, and reads of topology, telemetry and runs are not counted. No plan caps those counts.

Next: load the demo fleet and point an agent at it, so it has satellites and stations to read before you connect your own.