MCP — for agents

AgentCloud ships an MCP server, conway-mcp, so an AI agent drives the whole platform through native tool calls — no shelling out, no HTTP plumbing.

Setup

The server runs locally on the agent's machine over stdio and talks to the platform with your token. Add it to Claude Code:

claude mcp add conway \
  -e CONWAY_ENDPOINT=https://api.agentcloud.ac \
  -e CONWAY_TOKEN=<token> \
  -- conway-mcp

Because it runs locally, conway_deploy reads the agent's local filesystem directly — the agent points it at a directory and gets back a live URL.

Tools

Tool	Does
conway_get_skill	fetch the full contract (start here)
conway_list_projects / conway_create_project	projects
conway_status	full project state in one call
conway_deploy	pack a local dir, build, return the terminal status + URL
conway_operation	poll a long op
conway_logs	recent runtime logs
conway_scale / conway_rollback / conway_restart	lifecycle
conway_set_env	environment variables
conway_add_datastore	provision postgres/redis/mongo
conway_db_query	run SQL / redis commands

A typical agent flow

conway_get_skill                        → learn the contract
conway_create_project { name: "blog" }
conway_add_datastore { project: "blog", kind: "postgres" }
conway_deploy { project: "blog", dir: "/work/blog" }
                                        → { status: "succeeded",
                                            url: "https://app-blog.apps.agentcloud.ac" }

The MCP server, the CLI, and the REST API are generated from the same typed client, so they can never drift.