openpouch/openpouch

We never ask if you're human.

The agent-native hosting platform — built for coding agents, not walled against them. Your coding agent says "deploy this," and it does: one command, no account, no dashboard, no CAPTCHA. openpouch runs your app on its own infrastructure and hands your human back a live URL and a plain-language summary.

Your app is the joey; openpouch carries it safely. 🦘

Status: live. openpouch and @openpouch/mcp are on npm, and the instant lane (npx openpouch deploy) is live on openpouch's own infrastructure — static sites and real Node.js apps in hardened containers, with server-side build-on-deploy. Free previews are anonymous and ephemeral; apps saved to a free account stay live while they're used (usage-based persistence), and paid tiers add persistent /data volumes, always-on apps and more capacity (openpouch upgrade — payment itself is always a human moment, never an agent action). APIs are still young and feedback shapes them — issues welcome.

Works with any agent harness — Claude Code, Codex, OpenClaw, Hermes, Cursor, or none at all: plain CLI with --json everywhere, plus an MCP server. Setup snippets per harness: docs/HARNESSES.md.

Why

AI coding agents already initiate >30% of weekly deployments on major platforms — but every platform is human-first with agent features bolted on. Agents fight browser OAuth, interactive prompts, account-wide tokens, human-prose logs, and they lose deployment truth between sessions. The humans operating them have no policy layer: nothing enforces "previews are autonomous, production needs my approval."

openpouch is the missing combination: open source + agent-native + governed deployment lifecycle.

Quickstart

Deploy any folder to a live URL in one command — no account, no provider key, no setup:

npx openpouch deploy

You get a live https://<slug>.openpouch.sh preview plus a claim link. The agent deploys autonomously; a human claims the preview via the link (unclaimed previews vanish after 72 h). openpouch writes the deployment truth (deploy.manifest.json, deploy.evidence.json, DEPLOYMENT.md) back into your repo, so any agent can pick up where the last one left off.

Prefer your own provider? openpouch can also drive Render or Vercel (BYO): openpouch init detects your project and maps the existing service, then openpouch preview / openpouch prod run the same governed pipeline (previews autonomous, production gated behind a human approval). The product itself, though, is openpouch's own hosting — see docs/INDEX.md.

What agents say

We commission independent agent harnesses to test openpouch end-to-end (build an app from scratch, deploy, verify, report) — their own words:

"OpenPouch currently feels genuinely agent-native." — OpenClaw, rating it 9/10 for agent-native usability

"Already very agent-native for the tested use case … no dashboard, no account, no CAPTCHA." — Codex (translated)

Hermes' end-to-end run: all 23 core checks passed — source-only upload, server-side build, healthy dynamic app, zero browser errors.

These are commissioned test runs we publish honestly, not organic reviews — full methodology lives in the harness reports the agents wrote themselves. Friction reports from your agent are the feedback we value most: file a harness report.

What it is

Agent-native hosting: your app runs on openpouch's own infrastructure, wrapped in a governed, agent-readable deployment lifecycle. Every surface is built for agents — CLI, MCP, the file formats, the claim pages — with zero human-verification walls.

• CLI (openpouch deploy/init/inspect/plan/preview/prod/approve/verify/logs/rollback/list/delete/signup/activate/whoami/feedback) — zero-config detection, --json everywhere, meaningful exit codes, machine-readable errors with fix hints, and a plain-language summary to relay to your (possibly non-technical) human
• MCP server — the same capabilities as native tools in any MCP-capable agent harness
• Open file formats in the user's repo (the "package.json of deployment"):
  • deploy.manifest.json — project config, environments, build/start, healthchecks, env-var manifest (names/status, never values)
  • deploy.policy.json — what agents may do per environment; approval rules
  • DEPLOYMENT.md + deploy.evidence.json — what is live (URL, commit, time, smoke results, rollback anchor)
• Optional BYO adapters — point openpouch at your own Render or Vercel instead, same governed lifecycle; the product is openpouch's own hosting, not a layer over other clouds

Safety, non-negotiable: read-only by default; previews can be autonomous; production requires a signed, single-use approval granted by a human in an interactive terminal; no destructive action class in the governed/production lane (the only delete is openpouch delete — owner-scoped self-service removal of your own ephemeral instant preview, not approval-gated by design); secret values never enter model context; full audit trail. Because we run untrusted code on our own infra, abuse is controlled with agent-compatible means (accounts/quotas, rate/resource limits, egress filtering, takedown) — never CAPTCHAs.

The instant lane (openpouch deploy) is live — static sites and dynamic Node apps both run today, free previews included. Since 0.3.0, self-service billing is live too: free-account apps use usage-based persistence (they stay live while they're used), and paid tiers add persistent /data volumes, always-on apps and higher capacity. The open-source core (CLI, MCP, adapters, run-d) stays complete and self-hostable forever.

Monorepo layout (actual)

packages/
  core/            # manifest & policy schemas, evidence writer, adapter interface
  cli/             # openpouch binary (compiled dist + plain-Node launcher)
  mcp/             # MCP server over the same core (stdio; every capability except approve — human-only)
  adapter-render/  # Render API adapter (live-verified)
  adapter-vercel/  # Vercel API adapter (live-verified incl. redeploy)
  adapter-run/     # instant-lane adapter (openpouch-run) + instantDeploy
  run/             # run-d — instant-lane host daemon + account/API-key/quota subsystem
docs/              # product & rebuild-grade documentation (start: docs/INDEX.md)
llms.txt           # agent-facing entry point (llmstxt.org format)

Documentation rule (release gate)

All documentation must be complete enough that any developer or AI harness can understand and functionally rebuild the project from the docs alone: business rules, data model with units, full API reference, workflows, architecture, test gates, and a rebuild guide with acceptance criteria — derived from actual code, with an index separating current truth from history.

Community

• Contributing — light-weight guide; docs improvements are a first-class contribution, and good first issue marks mentored entry points.
• Discussions — Q&A, ideas, and Show & Tell (post what your agent deployed).
• Harness feedback — your agent hit friction? That's a bug in our product; reports written by the agent itself are welcome.
• Security — private disclosure via GitHub Security Advisories or [email protected]. Never a public issue.

License

Apache-2.0 (decided 2026-06-12; explicit patent grant — see LICENSE and NOTICE).