Agency Orchestrator MCP Server: Setup for Claude, Cursor and Other Apps

• Runs a multi-step AI workflow from a YAML file

  Runs the steps of a workflow file in order, each handled by an AI expert role through the provider you pick. It returns the final result, the token count and the folder where the full run is saved, or says which steps failed.

  Try: “Run workflows/launch-plan.yaml with the topic 'new pricing page'.”

  Runs code · run_workflow

• Writes a workflow from a one-line idea

  Asks an AI model to draft a workflow YAML from your description, picks roles from the agents folder, checks it and saves it to a workflows folder in your user directory.

  Try: “Compose a workflow that researches a market and drafts a report.”

  Write · compose_workflow

• Checks a workflow file without running it

  Reads a workflow file and reports whether it is valid, without calling any model.

  Try: “Is this workflow file valid?”

  Files · validate_workflow

• Shows the order a workflow would run in

  Reads a workflow file and shows its execution plan as a dependency graph, without running it.

  Try: “Show me the plan for launch-plan.yaml.”

  Files · plan_workflow

• Lists the workflow templates available

  Lists workflow files found in the workflows folder with their names and descriptions.

  Try: “What workflows do I have?”

  Files · list_workflows

• Lists the AI expert roles

  Lists the roles found in the agents folder, which you can use as steps in a workflow.

  Try: “Which expert roles can I use?”

  Files · list_roles

Watch out for

• Running a workflow spends model credits

  Each step is a model call, and workflows with video or image steps are billed per use. The result shows token and media counts after the run. github.com

• It can start your installed AI CLIs

  With a CLI provider the server launches a coding CLI on your machine. The Claude Code step runs with its tools turned off in an empty temp folder, and the Codex step uses a read-only sandbox. A step that uses Antigravity can opt in to skipping its permission prompts in the workflow file. Other CLIs run under their own settings. github.com

• Your task text goes to the provider

  Workflow inputs and step outputs are sent to the model you chose. Use a local Ollama model for private material. github.com

• It writes run archives to your home folder

  Results and generated workflows are saved under ~/.ao unless you set AO_OUTPUT_DIR or AO_HOME. github.com

• A workflow file does not run shell commands

  The 0.21.1 source defines steps as model calls, not shell steps. Image and video steps can start local programs such as ffmpeg. Only run workflow files you trust, since they choose the provider, model and roles. github.com

Security advisories