AI agent tools that touch your real systems
AI agent tools are only useful if the agent can run them without a human in the middle. AgentUI exposes its whole platform as one CLI where every command returns structured JSON — so Claude Code, Codex, Cursor or Aider can list projects, query your database, invoke a function and ship an app, then read the result and keep going.
One flat markdown file, no sign-up, nothing to install to read it.
> agent: list the projects
$ agentui --json project list
{ "total": 2, "projects": [ … ] }
> agent: run the nightly report
$ agentui --json functions invoke nightly
{ "status": "ok", "rows": 1482 }
> agent: ship it
$ agentui --json deploy
{ "deployed": true }What the agent actually gets
Not a chat wrapper: fifteen command groups it can call directly. Projects to sync and push, functions to invoke and read logs from, environments to promote, integrations to list and configure, secrets to request, and validate, build and deploy to ship the result.
| agentui --json project list | Enumerate projects with id, type and status |
|---|---|
| agentui --json project sync <id> | Pull a workflow to disk as plain files |
| agentui --json project diff | See what changed before committing to it |
| agentui --json project push --dry-run | Print the exact operations a push would send |
| agentui --json functions invoke <id> | Run a backend function and read the result |
| agentui --json functions logs <stepId> | Read execution logs for debugging |
| agentui --json integrations list | Discover what the account can already reach |
| agentui --json secrets list | See which credentials exist — never their values |
| agentui --json validate | Check the project before shipping it |
| agentui --json deploy | Ship a built project |
Tools that reach the systems you already run
An agent that can only write files is not much use. These tools read and write SQL Server, Postgres, MySQL and BigQuery, and reach Salesforce, HubSpot, Shopify, Gmail, Teams, WhatsApp, Google Sheets, Excel and Odoo. When the integration you need is not in the catalog, a backend function becomes a tool the agent can invoke like any other.
What an agent cannot do with them
This is the part worth reading before you hand an agent your production data. Secret values never leave the server, so an agent can use a credential without ever seeing it. Path traversal is rejected on every read, write and delete. A push is capped at 50 operations and 1 MB, and --dry-run prints the exact diff first. Cross-tenant requests return 404, not 403 — the agent cannot even confirm another company's resource exists.
Wire it into the agent you already use
There is no SDK to adopt. Install the package, let the agent shell out, and point it at the flat markdown reference so it knows the command surface without guessing.
npm install -g @agentuiai/cliOne flat markdown file, no sign-up, nothing to install to read it.