Skip to Content

Agent tools

Your agent builds apps on manh.run through MCP tools; this page lists each group of tools, the two tool sets by connection URL and the order an agent calls them in a build session.

manh.run writes no code itself. Your agent (Claude Code, ChatGPT, Gemini, Cursor…) connects to https://mcp.manh.run, reads the docs, writes files into app/, runs check and deploys. Each of those steps is a tool, much like the CLI commands of other frameworks.

Tool groups

GroupToolsWhat for
Projectslist_projects · create_project · get_projectList, create and inspect projects
Fileslist_files · read_files · write_files · delete_filesRead and write the app/ tree
checkcheckFind errors before deploying
deploydeployPut the UI and the API live together, then run the tests
Testingrun_acceptance · create_test_accounts · remove_test_accountsBusiness tests, test accounts per role
Checkpointscheckpoint · restore_checkpointSave and restore the file tree
Databasedb_query · db_execRead and write the app’s database
LogslogsRequests, console output, traffic, metrics
Secretsenv_set · env_listAPI keys and tokens of outside services
Feature blocksblocks · block_plan · block_addAdd tested features to the app
Croncron_jobs · cron_runs · cron_controlInspect and control scheduled jobs
Outside dataintegrations · integration_sqlThe owner’s Supabase
Imageslist_images · upload_imageThe owner’s own photos for the website
Docsframework_docsFramework docs for the agent

Every tool that works on a project takes a project argument: the project id from list_projects. The agent sees only the projects of the account it signed in with.

Two tool sets

manh.run has two MCP URLs. They differ in the tools that write real data or take secrets.

  • https://mcp.manh.run: the full set, for Claude Code, Cursor, Codex and custom connectors.
  • https://mcp.manh.run/directory: the set used by the ChatGPT app directory. It has no env_set, db_exec, integrations or integration_sql. write_files does not delete; deleting is its own tool, delete_files. deploy has no force.

With the smaller set you set secrets in the dashboard (the project’s Secrets page) and change data through the app’s API or the Database tab.

A build session

An agent usually calls the tools in this order. The results of write_files, check, deploy, get_project and create_project end with a Progress section and a next: line naming the next step, so weaker models reach a working app too.

1. framework_docs("process"), ("framework"), ("pro") read the conventions before writing 2. list_projects / create_project pick or create a project 3. write_files, one phase at a time tables + repository → handlers → layout + pages 4. check({ typecheck: true }) fix every error 5. deploy the API answers 503 until the first deploy; returns test results 6. run_acceptance fix until every test in spec/acceptance.json passes 7. remove_test_accounts, checkpoint clean up test accounts, save a labelled point

Good to know

  • Writing files never deploys. The live app changes only when the agent calls deploy or you click Deploy.
  • You see each file write right away in the dashboard preview.
  • Errors come back as ERROR_CODE: what to do, so the agent can fix them without guessing.

Next steps

Last updated