openpouch
OfficialThe openpouch server provides an agent-native deployment platform for autonomously deploying, managing, and monitoring web applications, with a governed lifecycle that integrates human approval for production deployments and critical actions.
Initialize a project: Auto-detect framework, build config, and env vars; generate
deploy.manifest.jsonanddeploy.policy.json(previews autonomous, production requires human approval by default).Instant zero-config deploy: Deploy any folder to a live
https://<slug>.openpouch.shURL with no account or provider key required, including server-side builds and secret env var injection.Deploy preview environment: Run the full governed pipeline (policy check → deploy → poll → smoke → evidence) for a preview environment — autonomous by default.
Deploy production environment: Run the governed pipeline for production. If policy requires human approval, returns an
approvalRequest{id}— agents intentionally cannot self-approve.Inspect deployment state: Read-only view of what is deployed, on which commit, any missing required env vars (names only, never values), and configuration drift.
Plan deployments: Read-only policy evaluation per environment — reports whether a deploy is allowed, requires approval, or is blocked, plus concrete next steps.
Verify a live deployment: Run health/smoke checks against a live URL and record results in evidence files.
Fetch runtime logs: Retrieve structured runtime logs (timestamp + message) for a deployed service to assist debugging.
Rollback a deployment: Redeploy the previously recorded rollback anchor commit, subject to governance policy and potential human approval.
Check account identity: Report the account behind the current API key — tier/plan, usage stats, or anonymous status. Key is never echoed.
List deployed apps: Enumerate instant-lane apps by name/slug, kind (static/dynamic), status, live URL, and expiry.
Delete an app: Remove an owned instant-lane app by slug to free a quota slot. No approval needed for ephemeral previews.
List pending approvals: Read-only view of pending human approval requests — approving is intentionally human-only via the interactive CLI.
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.
Related MCP server: shiply — Static Site Hosting & Deployment
Quickstart
Deploy any folder to a live URL in one command — no account, no provider key, no setup:
npx openpouch deployYou 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,--jsoneverywhere, meaningful exit codes, machine-readable errors with fix hints, and a plain-languagesummaryto relay to your (possibly non-technical) humanMCP 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 rulesDEPLOYMENT.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 issuemarks 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 security@openpouch.dev. Never a public issue.
License
Apache-2.0 (decided 2026-06-12; explicit patent grant — see LICENSE and NOTICE).
Available Tools
13 toolsopenpouch_deleteA
Delete one of YOUR OWN instant-lane apps by name (slug) → frees a quota slot. This is the self-service fix for a 'limit reached' deploy error: list your apps, delete an unused one, redeploy. Owner-only — it uses your account key and the server verifies ownership, so you can never delete another account's app. Deleting an ephemeral preview is NOT a governed-production action, so it needs no approval (there is still no approve tool, D13). Returns the removed name + a plain-language summary.
Inputs: slug (required).
openpouch_deployA
Zero-config INSTANT preview on openpouch's own infra — the openpouch deploy command (was CLI-only; Codex 2026-07-04). No account, no provider key, no manifest needed (a saved openpouch API key lifts the deploy into your tier; otherwise anonymous). Uploads the folder, builds on deploy (dynamic Node apps run in a container), probes health, and returns top-level url, healthStatus (+ pending), and a plain-language summary to relay to your human. The private claim link (a save token — like a password) is REDACTED from the result BY DEFAULT here, because tool results flow through chat context; it is saved locally to .openpouch/claim.json (0600, gitignored), so nothing is lost. Env var VALUES are secrets: injected into the container only, never in output or evidence (names only). For a full-stack app set healthPath (e.g. /api/health) so the deploy is held to its API too, not just /.
Inputs: cwd, dir, env, healthPath, +1 more.
openpouch_deploy_previewA
Deploy the preview environment through the governed pipeline (policy check → deploy → poll → smoke → evidence). Default policy allows previews autonomously. For the zero-config instant preview (no manifest/provider key) use openpouch_deploy instead. On success the live URL is returned at top-level url, plus a plain-language summary to relay to your human.
Inputs: cwd.
openpouch_deploy_productionA
Deploy production through the governed pipeline. If policy requires approval, this returns approvalRequest{id} with isError — ask your HUMAN to run openpouch approve <id> in their own terminal, then call this tool again. Agents cannot approve; there is intentionally no approve tool. Every result carries a plain-language summary to relay to your human; on a live deploy the URL is at top-level url.
Inputs: cwd.
openpouch_initA
Initialize a project for openpouch: detects framework/build/env vars, writes deploy.manifest.json + deploy.policy.json (default policy: previews autonomous, production requires human approval), auto-matches an existing provider service by name. Idempotent.
Inputs: cwd, force.
openpouch_inspectARead-only
Answer: what is deployed, where, on which commit, which required env vars are missing (names only — never values), and what drift exists between manifest and provider. Env vars passed at deploy time (--var/--env-file) appear as deployProvided (names from local evidence — the instant lane never exposes them via the API). Read-only. The result carries a plain-language summary you can relay directly to a non-technical human.
Inputs: cwd.
openpouch_listARead-only
List the instant-lane apps owned by the current openpouch account key: name (slug), kind (static/dynamic), status, live URL, and expiry. Read-only; needs an account key (anonymous previews aren't grouped under an account → an empty list with a signup pointer). Carries a plain-language summary to relay to your human. Pair with openpouch_delete to free a quota slot when a deploy hits a 'limit reached' error.
Inputs: no named parameters.
openpouch_list_approvalsARead-only
List pending approval requests (read-only). Approving itself is human-only via openpouch approve <id> in an interactive terminal — by design not available to agents.
Inputs: cwd.
openpouch_logsARead-only
Structured runtime logs ({timestamp, message}) of the mapped service. Read-only. This is the first step of the self-repair loop: when a deploy comes back unhealthy (a failed verify, a non-200 app, or a pending/not-yet-live URL), read the logs to find the cause, fix the code or config, redeploy, then verify — repeat until it passes.
Inputs: cwd, limit, environment.
openpouch_planARead-only
Per environment: the policy decision (allowed / requires-approval / denied), blockers, readiness, and concrete next steps including the human-approval path. Read-only — reports, never acts.
Inputs: cwd.
openpouch_rollbackA
Redeploy the recorded rollback anchor commit (the deploy that was live before the latest one). Approval-gated like any write action — may return approvalRequest{id} for your human.
Inputs: cwd, environment.
openpouch_verifyA
Run the healthcheck/smoke against the environment's live URL and append the results to the evidence (deploy.evidence.json + DEPLOYMENT.md). For a full-stack app pass healthPath (e.g. /api/health) — it is checked IN ADDITION to /, because a live shell can hide a dead API. The result carries a plain-language summary (healthy / has a problem) you can relay to your human.
Inputs: cwd, healthPath, environment.
openpouch_whoamiARead-only
Report the openpouch account behind the current API key: tier (plan) and current usage (live apps, deploys this hour/day), or that the caller is anonymous. Read-only. The key is read from OPENPOUCH_API_KEY / ~/.openpouch/openpouch-run.key and never echoed. Carries a plain-language summary you can relay to your human. (Signup/activation are CLI/web only — they involve a human email or browser step.)
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 13 tools
Each tool has a distinct purpose clearly described: deploy, delete, list, inspect, logs, verify, rollback, init, plan, whoami, approvals. No two tools overlap significantly; even deploy variants are differentiated by pipeline type.
All tools follow the pattern 'openpouch_<verb>' or 'openpouch_<verb>_<noun>' using snake_case consistently. Examples: openpouch_delete, openpouch_deploy_preview, openpouch_list_approvals. No mixing of conventions.
13 tools is well-scoped for a deployment management server. It covers initialization, multiple deploy types, listing, inspection, logs, verification, rollback, deletion, planning, and account info without being excessive.
The tool set covers the core app lifecycle: init, deploy, inspect, logs, verify, rollback, delete. Missing direct env var management (setting/removing) but env vars are handled during deploy and inspect shows missing ones. Approvals are intentionally not automated, so not a gap.
Maintenance
Related MCP Connectors
Deploy AI-generated HTML/CSS/JS to instant public HTTPS URLs from any MCP-compatible agent.
Hosting for AI agents: your AI client deploys Docker apps to live HTTPS URLs over MCP.
Hosted MCP for creating, checking, deploying, and hosting static sites for AI agents.
Hosting for AI agents: publish a live website in one tool call, ephemeral or forever.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceMCP server for Deloc — deploy static sites from AI coding agents. Say "deploy this to Deloc" in Claude Code or Cursor and get a live URL back without leaving your editor.4MIT
- AlicenseAqualityCmaintenanceInstant web hosting for AI agents. Publish a live site in one call, no account needed.5MIT
- AlicenseNot gradedqualityDmaintenanceMCP server for self-hosted static site publishing on Cloudflare Workers. Enables AI coding agents to deploy pages with a single 'publish' tool and get live URLs, with support for atomic updates, versioning, and per-site passwords.8 npm2MIT

agenthost MCP serverofficial
FlicenseNot gradedqualityDmaintenanceAgent-first web hosting driven entirely through a remote MCP server, enabling full hosting management via natural language with OAuth or token authentication.-