Buildfyio for AI agents
Buildfyio ships an official @buildfyio/mcp server so agents that speak the Model Context Protocol — Claude Code, Claude Desktop, Cursor, Windsurf — can create projects, deploy from git, tail logs, and roll back releases in plain English.
One-command install
npm i -g @buildfyio/cli buildfy login buildfy mcp install claude-code # or claude-desktop | cursor | windsurf
The CLI mints a scoped API token (default: read + deploy), writes the client's mcp_servers.json, and asks you to restart the host. That's it.
Tools
| list_projects | read | List projects in an org. |
| create_project | write | Create a project from a git URL. |
| deploy | deploy | Trigger a deployment. |
| get_deployment_status | read | Status, URL, failure reason. |
| get_logs | read | build / deploy / runtime logs. |
| get_build_failure | read | Structured JSON diagnosis of a failed build — the agent-native version of get_logs. |
| set_env_var | write | Create or update an environment variable. |
| rollback | write | Roll production back to a previous deploy. Destructive — requires confirm: true. |
Scoping the token
Every tool is guarded by a scope on the API token. If a read-only token calls deploy, the server refuses with a TOKEN_MISSING_SCOPE error the agent can read and explain. Change scopes with:
buildfy mcp install claude-code --scopes read buildfy mcp install claude-code --scopes read,deploy # default buildfy mcp install claude-code --scopes read,deploy,write # full access
Audit
Anything an MCP client does is audited with actor_type: ai_agent. Filter the Audit tab in your dashboard to see exactly what your agent touched, when.
Remote server
If a stdio install isn't practical (Claude.ai web, custom agent), point at the hosted endpoint:
https://mcp.buildfyio.com/v1/mcp OAuth 2.1 + PKCE. authorization_endpoint = https://buildfyio.com/mcp/authorize scopes = read | deploy | write
Eval status
The golden-task suite runs nightly against Claude Sonnet — pass rate is published on the status page. The target is >= 90%.