Deploy from Claude Code, Cursor and Codex
The Inquir CLI is also an MCP server. Connect it to your agent, and the agent ships the code, calls the service, reads build errors and logs, fixes the code and ships again. You look at the result instead of copying commands back and forth.
Last updated: 2026-10-18
- claude mcp add inquir -- inquir mcp, and the tools are in Claude Code
- Node.js, Python and Go: handlers and HTTP servers, no Dockerfile needed
- Rolling back to the previous deploy is one more tool call
No card required, then Simple at $5/month.
The gap
The agent wrote the code. The rest is on you
Claude Code, Cursor and Codex write services with confidence and then stop at the deploy: a server, a Dockerfile, variables, a domain. You copy commands, paste error logs back into the chat and wait for the next round.
The longer that loop runs through you, the slower the agent gets: it cannot see what broke on the server, so it fixes things blind.
How it works
Inquir as an MCP server
inquir mcp starts an MCP server over stdio and reuses your CLI login. The agent gets tools to deploy, invoke, run commands, read deploy status and invocation history, and roll back. The tools work with functions and web functions; container applications and databases the agent creates with the CLI in a terminal (inquir apps …).
deploy takes a folder: it detects the runtime (Node.js, Python, Go) and the mode, either a handler or an HTTP server started with PORT set (for a Go server, pass the start command). Dependencies install on the server; when a Node.js or Python function needs system packages such as ffmpeg or Chromium, the agent lists them in inquir.yaml.
After a deploy the agent reads the build status with its error text and the invocation history with logs, and fixes the code from real output rather than guesses.
Setup
Connect in a minute
Install the CLI and log in once; the agent uses that login from then on. The CLI needs Node.js 22 or newer.
npm install -g @inquir/compute-cli@latest inquir login # Claude Code claude mcp add inquir -- inquir mcp
{ "mcpServers": { "inquir": { "command": "inquir", "args": ["mcp"] } } }
[mcp_servers.inquir] command = "inquir" args = ["mcp"]
The loop
The agent’s loop
Writes the code
The agent builds a service in Node.js, Python or Go: a handler or an ordinary HTTP server.
Describes the build
When a Node.js or Python service needs system packages such as ffmpeg, the agent adds inquir.yaml.
Deploys
It calls deploy with the folder path. The build runs on the server; no local Docker.
Checks
It calls the service with invoke and reads deploy_status and list_invocations.
Fixes
It edits the code from the real logs and deploys again; if things got worse, rollback.
MCP tools
The tools the agent gets
deploy
Deploy a folder as a function or a web service and get the build result, errors included.
invoke
Call a function or a web service with a JSON payload and get the result and the logs.
list_invocations
Recent invocations with status, duration, errors and logs: the starting point for debugging.
deploy_status
The state of one deploy: queued, building, deploying, succeeded or failed, with details.
rollback
Return a function to a previous successful deploy.
exec
Run a shell command against a private copy of the function source or, read-only, against the deployed release.
list_functions
What is already deployed in the workspace, with runtime, mode and status.
Safety
Why giving the agent SSH is a bad idea
SSH access hands the agent everything at once: deleting data, changing configuration, opening ports. Untangling its mistake is then your job.
A PaaS built around a web console is awkward for an agent: it needs tools with a clear input and output, not buttons.
What works is a narrow set of actions (deploy, invoke, read logs, roll back) with the same permissions you have and within your plan’s limits.
When it fits
When it fits
When this works
- You build services with an agent and want it to take them all the way to a working URL
- Bots, APIs, webhooks and background jobs without infrastructure of your own
- You want a clear rollback for when the agent gets it wrong
When to skip it
- The agent must manage arbitrary infrastructure such as servers, networks or DNS: the Inquir MCP server does not do that
- Unreviewed changes to production with critical data: give the agent a separate workspace instead
FAQ
FAQ
What can the agent do with my account?
What you can: the MCP server uses the token that inquir login stored, and that token reaches every workspace you belong to. Making a separate workspace active with inquir use keeps the agent’s work apart by convention, not as a hard boundary. For a hard boundary, run the agent as a separate user invited only to that workspace with a limited role.
Do I need Docker?
No. deploy packs the folder and the platform installs dependencies and builds it on the server.
Does it work with Codex and other clients?
inquir mcp is a regular MCP server over stdio, so any client that can start a local command works: Claude Code, Cursor, Codex and others. The Codex configuration is in the example above.
What if the agent deploys broken code?
The agent sees the error in deploy_status and in the invocation logs. Going back to the last good version is one rollback call; you can do the same in the dashboard or with inquir rollback.
Can I do without MCP?
Yes. An agent with a terminal can call the CLI itself: inquir deploy, inquir logs, inquir rollback. The CLI package ships a skill file with instructions for agents.
What does it cost?
The agent uses your plan’s resources: Simple is $5 and Pro $20 per month (590 ₽ and 2 490 ₽ in Russia). The plan’s CPU and memory limits apply to everything the agent deploys.