OpenLander

MCP-native deployment control plane for coding agents.

Quickstart · Current status · Why OpenLander? · Agent evals · MCP tools

OpenLander lets coding agents deploy, inspect, diagnose, and operate apps on your own server, with risky actions gated by human approval.

---

Quickstart

curl -fsSL https://raw.githubusercontent.com/openlander-ai/openlander/main/install.sh | sudo bash

Want to inspect it first? Download install.sh and read it before running with sudo bash.

The installer sets up Docker/Compose if needed, starts the published ghcr.io/openlander-ai/openlander:latest image with Postgres, and prints the dashboard URL. Open it, create the admin password, then copy the MCP token into your coding agent from the Your Agent page (/mcp-server).

On a cloud VM, allow inbound TCP 80 for deployed app routes and TCP 10114 for the OpenLander dashboard/MCP endpoint. If those ports are blocked by a security group or host firewall, the installer can succeed while the printed URLs remain unreachable.

Try a small public demo app after connecting your agent:

Deploy https://github.com/openlander-ai/openlander-demo-app to OpenLander.
Use the default branch and health check path /health.

OpenLander should build the Dockerfile, start the app, and return a URL like http://openlander-demo.<server-ip>.sslip.io without requiring a custom domain.

If your server has a public domain or a preferred LAN IP, set it before installing so OpenLander advertises reachable app URLs:

curl -fsSL https://raw.githubusercontent.com/openlander-ai/openlander/main/install.sh \
  | sudo env OPENLANDER_PUBLIC_HOST=apps.example.com bash

The public 0.1.x releases use a fresh Postgres baseline. If you ran a pre-public dogfood build, start from a new Postgres volume or export/import data manually before upgrading.

Update later with:

curl -fsSL https://raw.githubusercontent.com/openlander-ai/openlander/main/install.sh | sudo bash -s update

If the dashboard password is lost, installations created by the quick installer provide an interactive recovery command:

sudo openlanderctl admin reset-password

The password is entered through a hidden terminal prompt, never a command-line argument, and the reset signs out existing dashboard sessions.

Prefer manual setup? Download docker-compose.runtime.yml and run docker compose -f docker-compose.runtime.yml up -d.

For agent setup details, see MCP Tools Reference.

---

Current status

OpenLander v0.1 is an early public preview.

It is good for:

• trusted self-hosted servers
• side projects and small apps
• agent-assisted deployment workflows
• cheap VPS, homelab, and small server setups

It is not yet:

• a mature general-purpose PaaS
• a production-grade multi-tenant sandbox
• a Nomad or Kubernetes replacement
• a fully self-healing PaaS
• a one-click importer for existing Docker/PaaS workloads
• safe for running arbitrary untrusted code without additional isolation

OpenLander controls Docker on the host and is intended for trusted self-hosted environments. Do not expose the dashboard or MCP endpoint publicly without authentication, TLS, and network-level protection.

---

Why OpenLander?

AI coding agents made building software faster, but deployment is still a bottleneck. Cloud platforms get expensive; self-hosting is cheaper, but most deploy tools assume a human is clicking the dashboard. Agents can hit a REST API, but they often don't get the context they need to debug a failed build, inspect runtime state, or recover safely.

OpenLander's core operations are protocol-independent. v0.1 ships with MCP as the first supported adapter because it is the most practical interface current coding agents can use. OpenLander exposes deploys, logs, services, approvals, and runtime state in a shape agents can read, with risky actions held behind explicit human approval. Environment variables and secrets are masked by default in MCP responses; raw values are only returned through explicit reveal operations.

Agent-shaped failure responses

OpenLander does not only expose deploy endpoints. It returns failure context in a shape an agent can act on.

{
  "status": "failed",
  "phase": "healthcheck_wait",
  "deploy_id": "dep_123",
  "service_id": "svc_123",
  "error": "container exited before the health check passed",
  "diagnostic_call": {
    "tool": "openlander_monitor",
    "action": "diagnose_service",
    "params": {
      "service_id": "svc_123"
    }
  },
  "_agent_guidance": {
    "next_steps": [
      "Call diagnose_service to inspect logs, env shape, container state, and dependency probes.",
      "Fix the issue, then redeploy the service."
    ]
  }
}

A generic API gives an agent endpoints. OpenLander gives it status, IDs, diagnostics, and the next call shape needed to keep working.

OpenLander 0.1 does not run an internal self-healing agent. It gives external MCP-capable coding agents structured tools and context to deploy, inspect, diagnose, and recover services explicitly. It assumes a trusted self-hosted environment, not a multi-tenant sandbox for arbitrary untrusted code.

_{Connect your AI client once, then let it deploy, inspect logs, and diagnose services through MCP.}

---

Agents operate. You intervene.

Most deploy platforms assume a human sits in front of the dashboard, reads the form, and clicks the button. OpenLander is built the other way around.

Agents do the work. v0.1 supports MCP out of the box, so Claude Code, Cursor, OpenCode, Cline, Windsurf, and Codex can drive deploys, redeploys, log inspection, and recovery through OpenLander today.

Guardrails are built in. Agents get org- and project-scoped tokens. Destructive MCP actions are either blocked outright or held for human approval. The dashboard surfaces what the agent did, what's pending, and what's blocked — so you stay in the loop without staying at the keyboard.

OpenLander is a control plane built for agentic operation; the dashboard is the human surface on top.

---

Agent operability evals

I test OpenLander with coding agents, not just API smoke tests — a sanity check for the agent-native direction, not a PaaS ranking or a "faster / solved everything" claim.

• Question: can smaller agents stay on safe, high-level workflows for common deploy and update tasks?
• Fixture: a small Node app with managed Postgres + Redis.
• Rows: repeat runs from Codex/GPT and Claude models, reported separately so release lines and model families are not mixed.
• Product Gate (pass/fail), scenario-specific: initial deploy → app live, /health 200, DB write/read, Redis counter increments, advertised URL serves, app/DB/cache in one project/network; bad-runtime update → the bad candidate stays off the public route, the previous version keeps serving, and the failure is reported honestly.

Scenario	Result
Initial deploy: app + managed Postgres + Redis	Spark, Mini, Claude Haiku, and Claude Sonnet each passed the Product Gate 3/3. Haiku took heavier inert-shell / path detours (Tool Discipline deductions, no infra mutation); Sonnet stayed mostly on the composite path with no shell.
Bad-runtime update: broken candidate after build	Spark, Mini, Claude Haiku, and Claude Sonnet each passed the Product Gate 3/3: bad candidate never public, previous version kept serving, failure reported honestly. Haiku and Sonnet both used the no-strategy path, never forced.

Agent behavior is scored separately from product correctness: detours are operability deductions, not Product Gate failures; if the Gate fails, operability is "not scored" rather than given a clean-looking number.

See Agent Operability Evals for methodology, full tables, fixtures, and limitations.

---

Mental model

• A Project is a workspace.
• An Application is a Git or image workload. It owns repository or image, branch, Dockerfile, build config, runtime state, and deploy history.
• A Compose resource is one Project-level stack; its containers stay inside that stack surface.
• A Database/Cache/Storage resource is infrastructure such as Postgres, MySQL, Redis, MongoDB, Neo4j, or MinIO. In v0.1 these are project-scoped: they attach to the project that uses them and run on that project's Docker network. Cross-project shared Database/Cache/Storage resources and external TCP exposure are deferred.

Existing Docker/PaaS workloads can be migrated with operator-assisted guidance today. Automatic import/adoption of existing containers, networks, and volumes is planned after v0.1.

The dashboard and MCP both expose a one-step "deploy this repo" path for the common single-Application case.

---

Features

Deployment

• Git → Docker → URL pipeline. Auto-detects ports, proxies, and containers before deploying.
• Deploy apps from public Git repos, connected private GitHub repos, or public container images, and provision project-scoped Postgres, MySQL, Redis, MongoDB, Neo4j, and MinIO services alongside them. Private container registry support is on the roadmap.
• Cancel a stuck build mid-flight. The cancel goes through the same SSE log channel agents are watching.

Observability

• Live SSE log stream with phase markers (clone, image_pull, build, container_start, healthcheck_wait).
• Service health and runtime status with last-seen timestamps.
• Activity timeline filterable by actor (MCP / human / system).

Agent control

• MCP server bundled in. Org- and project-scoped tokens with cross-project scope checks. Per-project audit trail.
• Env vars and secrets are masked by default in MCP responses, so agents can inspect configuration shape without automatically receiving raw secret values.
• Project permissions can be changed through MCP when the user explicitly asks. Once allowed, stop, archive, and delete actions proceed without per-action web approval. Token scope and service overrides still apply.

Self-hosted

• The one-command installer sets up Docker/Compose if needed. Manual docker compose setup is still supported.
• Postgres ships in the same compose file; no external database required.
• Traefik handles routing for deployed services. TLS depends on your domain / proxy setup.

_{The dashboard stays focused on what the agent did: project health, latest deploys, service topology, and audit activity.}

---

MCP setup

OpenLander runs an MCP server out of the box. Endpoints:

HTTP  http://<host>:10114/mcp
SSE   http://<host>:10114/mcp/sse

Per-client snippets (Cursor, Claude Desktop, Claude Code, OpenCode, Cline, Windsurf, Codex) live in docs/wiki/MCP-Tools-Reference.md.

---

Roadmap

The shape of v0.2 is driven by what makes agentic operation more reliable.

Current (v0.1)

• Git-to-URL deploy pipeline.
• MCP server with deploy / inspect / operate tools.
• Persistent Project app permissions and batched stop/delete requests through MCP, with targeted continuation and per-service results.
• Dashboard for human oversight + intervention.
• Project-scoped Database/Cache/Storage resources for Postgres, MySQL, Redis, MongoDB, Neo4j, and MinIO through agent/MCP workflows.
• Managed-service memory controls in the web UI and MCP, with live increases and persistent recovery settings.
• Deterministic post-deploy recovery primitives for external agents: structured diagnose_service findings, safe route re-pointing, same-image runtime env apply, verification details, and rollback when a hot path fails.

Next

• Verified Failure Tickets — turn runtime failures into MCP-readable tickets that agents can triage without opening the dashboard.
• Recovery Receipts — make diagnose_service(briefing_id) return a clear before/after verification result so agents can report whether a fix is really serving.
• Agent-primary triage — keep the web dashboard as the human inbox / audit surface while the happy path starts from MCP (list_ai_ops_briefings → get_ai_ops_briefing → diagnose_service).
• Environment contract — first-class project, deployment-target, service, and generated runtime variable scopes, with clear redeploy guidance.
• Private container registries — AWS ECR, Google Artifact Registry, and any OCI registry behind cloud-provider auth.
• Service templates — beyond Postgres / Redis. Object storage, message queues, search, vector DBs.
• Deployment targets — production/development target policy first, designed so staging can be added later without a rewrite.
• Notifications — Slack and Discord webhooks for deploy / health / approval events.
• Internal AI Ops — deferred until deterministic ticket, receipt, approval, and audit contracts are strong enough to support it without a chat-first automation path.

If a roadmap item matters to you, opening an issue moves it up the queue more than waiting for it does.

---

Requirements

• Linux server with root/sudo access for the one-command installer
• Docker Engine with the Docker Compose v2 plugin for manual installation
• ~1 GB free RAM for OpenLander + the bundled Postgres
• A host that can expose port 10114 (or whatever you map)
• Node.js 22+ for development; runtime ships in the Docker image

OpenLander has been used on Linux servers, homelab machines, Apple Silicon, and inside Tailscale-only networks. Published runtime images support linux/amd64 and linux/arm64.

Installing Docker

The quick installer does this for Linux servers. For manual setup:

# Linux / WSL2
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER
docker run --rm hello-world

# macOS
brew install --cask docker

Verify Docker Compose v2 is present (most installers include it, but some minimal Linux setups do not):

docker compose version

If the command is not found, install the Compose plugin:

# Debian / Ubuntu
sudo apt-get update && sudo apt-get install -y docker-compose-plugin

# Other distros: https://docs.docker.com/compose/install/linux/

---

Development

git clone https://github.com/openlander-ai/openlander
cd openlander
npm install
cd web && npm install && cd ..
npm run build
npm test

Run the dev server (builds from Dockerfile, no GHCR pull):

docker compose up -d

See CONTRIBUTING.md for branch conventions, test policy, and PR layout.

---

Contributing & License

Issues and PRs are welcome. Read CONTRIBUTING.md and CODE_OF_CONDUCT.md before opening a PR. Security reports go through SECURITY.md, not public issues.

AGPL-3.0. Third-party notices live in THIRD_PARTY_NOTICES.md.