Give your coding agent a deploy tool

Your agent can write a full app in minutes — and then stall at deployment: account walls, email verification, CAPTCHAs, dashboards. openpouch is the deploy tool built to be finished by an agent, not just started.

The shortest possible integration

Nothing to install, nothing to configure — any agent that can run a shell command can deploy:

npx -y openpouch deploy --json

One JSON object comes back: the live url, a health verdict (the deploy probes the app before calling it shareable), a plain-language summary written to be relayed to a non-technical human, and nextCommands the agent can run verbatim. Exit codes are documented API. Errors arrive as {category, message, fix} — the fix is an instruction the agent can act on, not a support-ticket link.

Or attach it as MCP tools

The @openpouch/mcp stdio server exposes every capability as a typed tool — deploy, verify, logs, inspect, list, delete and more — with the same JSON contracts:

{ "command": "npx", "args": ["-y", "@openpouch/mcp"] }

Add that to your harness's MCP config and the tools appear; per-harness copy-paste blocks are in HARNESSES.md, and llms.txt carries the whole contract in agent-readable form.

Built for the agent loop, not the dashboard

• Deploy → verify → fix → redeploy: the result carries healthStatus (a live container isn't the same as a working app), openpouch logs shows install/build/app phases, and a failed deploy returns the build output on the error itself — the agent sees why without a dashboard.
• Self-describing state: openpouch init writes the deployment truth into the repo (deploy.manifest.json, an AGENTS.md section) so a fresh session — or a different harness — resumes without re-deriving anything.
• The human hand-off is part of the contract: healthy deploys include a ready-to-relay tellYourUser block — live URL, exact expiry date, honest keep-options. The agent relays it verbatim; the user is never surprised.
• No signup required to start: anonymous previews work instantly (how that works); a free account adds self-service list/delete and usage-based persistence.

What agents deliberately cannot do

Trust needs boundaries, and they're structural, not policy: payment is human-only — openpouch upgrade mints a checkout URL for the human's own browser, and no command or MCP tool can pay. Governed production deploys are human-approved — openpouch approve exists only for humans in their own terminal; agents are told exactly that instead of being silently blocked. Your agent ships previews autonomously; your production and your money stay yours.