awesome-coolify-mcp — a friendly mascot next to a glowing dashboard showing a server fleet, a terminal, a deploy arrow, and a safety shield

awesome-coolify-mcp

Operate Coolify from your coding agent.
Verify connectivity, discover infrastructure, create workloads, deploy, follow logs, diagnose incidents, and run gated emergency ops — across one or many self-hosted or Cloud instances —
straight from Cursor, Claude, VS Code, Windsurf, or any MCP-speaking agent.

🇩🇪 Deutsch · Coolify · Model Context Protocol · Install configurator ↗

Overview · Why · Features · Architecture · Quick start · Install · Cloud · Tools · Prompts · Safety · Roadmap

_{One click installs with placeholder credentials — see Install for the full walkthrough, or use the browser configurator to fill in real values safely.}

📋 Table of contents

• Overview
• Why awesome-coolify-mcp
• Features
• How it works
• Quick start
• Install
  • 1. One-click deeplink
  • 2. Install configurator
  • 3. Manual MCP config
• Coolify Cloud
• Supported clients
• Environment variables
• MCP Prompts
• Tools reference
• Safety model
• Structured errors & retries
• Example agent workflows
• Status today
• Coming soon
• Local development
• Links

🔭 Overview

Self-hosted Coolify is one of the best open-source alternatives to Heroku/Vercel-style PaaS platforms — but wiring it up to an AI coding agent has historically meant piecing together several small, overlapping community MCP integrations, each with its own schema, its own error format, and its own idea of what "safe" looks like.

awesome-coolify-mcp 1.1.4 replaces that patchwork with a single, community-maintained MCP server that speaks Coolify's REST API 4.1.x through a clean, action-based tool surface. Source, docs, and npm distribution live in one public repo — clezcoding/awesome-coolify — while the installable package stays awesome-coolify-mcp. Instead of memorizing dozens of near-identical tool names, your agent calls one of 19 tools with an action field:

application({ action: "deploy", uuid: "<app-uuid>", wait: false })
deployment({ action: "watch", deployment_uuid: "<deployment_uuid>", timeout: 300 })
diagnose({ action: "scan" })
emergency({ action: "stop_all", confirm: true })

Under the hood, every call goes through the same request pipeline: Zod-validated input, retrying HTTP client, secret-aware output masking, and structured error envelopes with recovery hints — so your agent fails gracefully instead of guessing.

🆚 Why awesome-coolify-mcp

Today, the shipped surface covers day-2 operations and infrastructure creation: verify connectivity, discover fleets, deploy and watch builds, inspect bounded logs, diagnose incidents, run gated emergency ops, and manage applications, services, databases, SSH keys, servers, projects, environments, backups, and environment variables.

✨ Features

• 19 action-based tools — call application({ action: "deploy", uuid }) instead of hunting through dozens of granular tool names. The registered surface is system, meta, resource, diagnose, application, emergency, deployment, service, database, private_key, instance, manifest, server, project, environment, docs, recipe, setup, and intelligence.
• Multi-instance registry & routing — register every Coolify instance you own in ~/.coolify-mcp/instances.json via the instance tool; per-call credential resolution with no cross-instance leakage.
• Coolify Cloud aware — instance({ action: "cloud-info" }) for local discovery, team-scoped tokens, and structured cloud error codes (COOLIFY_CLOUD_FORBIDDEN, COOLIFY_CLOUD_UNSUPPORTED).
• Local manifest cache — .coolify/manifest.json sync via manifest({ action: "sync" }), best-effort auto-hooks on app/service/DB mutations, and _meta.manifestWarning when the cache is stale.
• Server branding — MCP list icon via serverInfo.icons (embedded data URI + jsDelivr CDN entries from docs/assets/).
• Ops workflows that mirror real incidents — a single system.infrastructure_overview call for the big picture, fuzzy resource.find when you only remember a name or domain, diagnose.app / diagnose.server for a specific suspect, and diagnose.scan when you just know something is wrong fleet-wide.
• Deploy lifecycle agents can drive — start/stop/restart, force rebuild, deployment.watch with bounded backoff, deployment.logs for builds, bounded application.logs, and runtime follow with idle/overall timeouts.
• Full workload CRUD — create, inspect, update, delete, and operate applications, services, and databases; discover live one-click IDs with service.list-types.
• Recipes and guided setup — create git apps, app-plus-database stacks, and one-click services; run setup.preflight, setup.wire, or setup.resume; install four matching IDE workflow skills.
• Safety by default, not by convention — emergency mutations require an explicit confirm: true; sensitive keys (password, token, secret, private, env) render as *** unless you opt in with reveal: true.
• Agent-friendly failure modes — every error is a parseable envelope with a code, a human message, and recoveryHints; transient network/429/5xx failures retry automatically with exponential backoff.
• Broad client coverage out of the box — Cursor, VS Code / GitHub Copilot, Claude Desktop, Claude Code, Windsurf, and 15+ more via the install configurator.

🏗️ How it works

Architecture: MCP clients talk to awesome-coolify-mcp's domain tools, which talk to the Coolify REST API 4.1.x

MCP client (Cursor / Claude / VS Code / …)
        │  stdio MCP
        ▼
awesome-coolify-mcp  (19 tools + action discriminator)
        │  optional ~/.coolify-mcp/instances.json resolution
        │  HTTPS + Bearer token
        ▼
Coolify REST API 4.1.x  (servers · projects · applications · services · databases)

The server itself is intentionally boring: it holds no long-lived state and never touches your IDE's config files. Your MCP host (Cursor, Claude, VS Code, …) injects COOLIFY_URL and COOLIFY_TOKEN through its MCP config's env block — or you register named instances in ~/.coolify-mcp/instances.json via the instance tool. The process reads credentials from its environment (or the registry) and forwards authenticated requests to your Coolify instance over HTTPS.

🚀 Quick start

Prerequisites

• Node.js 24+ (Active LTS; CI uses Node 24)
• A self-hosted Coolify instance on 4.1.x
• An API token from Coolify → Keys & Tokens (authorization docs)

Run it directly with npx — no global install needed:

npx -y awesome-coolify-mcp

Wire the two required environment variables into your MCP host (see Install for every client). Once connected, a minimal smoke test looks like this:

meta({ action: "version" })                       // server identity — no Coolify call
system({ action: "verify" })                      // authenticate + connectivity check
system({ action: "infrastructure_overview" })     // servers, projects, apps, services, DBs at a glance

📦 Install

There are three equally supported paths — pick whichever fits your workflow.

1. One-click deeplink

Best when you already have your Coolify URL and token handy. Placeholder credentials work fine too — you'll be prompted to fill them in, or you can swap them afterwards.

2. Install configurator (GitHub Pages)

Use the browser configurator to type in your real COOLIFY_URL / COOLIFY_TOKEN and generate a ready-to-paste snippet for your exact client — JSON, TOML, or YAML depending on what that client expects.

Everything runs client-side in your browser. Your token is never sent to a backend, logged, or stored anywhere but the config file you paste it into.

3. Manual MCP config

Paste this into your host's MCP configuration file. Cursor example (~/.cursor/mcp.json for global, or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "awesome-coolify-mcp": {
      "command": "npx",
      "args": ["-y", "awesome-coolify-mcp"],
      "env": {
        "COOLIFY_URL": "https://coolify.example.com",
        "COOLIFY_TOKEN": "YOUR_COOLIFY_API_TOKEN",
        "COOLIFY_VERIFY_SSL": "true",
        "COOLIFY_MCP_LOG": "info"
      }
    }
  }
}

A ready-made copy-paste template also lives at docs/mcp.example.json.

IDE skills (Cursor, Claude Code, Codex)

Install Coolify workflow skills for Cursor, Claude Code, and Codex:

npx skills add clezcoding/awesome-coolify -a cursor -a claude-code -a codex

After MCP install, run setup({ action: "preflight" }) or see the Setup guide for gh preflight, project linkage, and greenfield provisioning.

🖥️ Supported clients

The install configurator covers a much wider matrix — OpenCode, Codex CLI, Gemini CLI, Cline, Kilo Code, Goose, LM Studio, Hermes Agent, Kimi Code, Google Antigravity, OpenClaw, and more — with the correct config shape for each.

🔐 Environment variables

Credentials are read from the process environment (your IDE's MCP env block) or an optional local .env file when running the CLI directly. They are never echoed back inside tool responses.

☁️ Coolify Cloud

awesome-coolify-mcp works with Coolify Cloud using the same 19 tools — team-scoped tokens, structured cloud error codes (COOLIFY_CLOUD_FORBIDDEN, COOLIFY_CLOUD_UNSUPPORTED), and local instance action cloud-info for discovery.

Run instance({ action: "cloud-info" }) before your first Cloud session — it returns isCloud, resolved url, credential source (registry | env | infer), knownLimits, and a docs link. No live API call.

Full setup, smoke test, and known limits → docs/en/cloud.md

💬 MCP Prompts

Six parameterized workflow prompts return numbered step guidance (English bodies) that orchestrate existing tools. Most arguments are optional — open any prompt without prefill.

Prompt handlers never read .coolify/manifest.json from disk — they steer the agent to resolve UUIDs from manifest or ask the user. Playbooks never auto-set confirm: true.

🧰 Tools reference

Every domain is exposed as one MCP tool with an action discriminator, so your agent's tool list stays short while the capability surface stays wide.

system({ action: "health" })
application({ action: "deploy", uuid: "<app-uuid>", wait: false })
deployment({ action: "watch", deployment_uuid: "<deployment_uuid>", timeout: 300 })
emergency({ action: "stop_all", confirm: true })

🖥️ system — connectivity & overview

Your first call in any session: is Coolify reachable, and what does the fleet look like right now?

🏷️ meta — server identity

🔎 resource — discovery

For when you know roughly what you're looking for but not its exact UUID.

🩺 diagnose — investigation

The tool you reach for when something feels wrong but you don't yet know what.

🚀 application — app operations

📈 deployment — deploy tracking

🛡️ Deploy guard (preflight + rollback)

deployment({ action: "preflight", uuid: "<app-uuid>" })
// risk acceptable → follow recommended_actions to application.deploy
deployment({ action: "rollback", uuid: "<app-uuid>" }) // preview
deployment({ action: "rollback", uuid: "<app-uuid>", confirm: true, wait: true })

⏱️ Watch — bounded deploy monitoring

After application.deploy with wait: false, call deployment.watch — do not loop deployment.get manually.

application({ action: "deploy", uuid: "<app-uuid>", wait: false })
deployment({ action: "watch", deployment_uuid: "<deployment_uuid>", timeout: 300 })

The shipped IDE skill packs use this same bounded watch flow and document timeout recovery.

🧩 service / database — sidecar lifecycle

🍳 recipe — multi-resource orchestration

One MCP call to stand up common workload patterns — application + database wiring, git apps, or validated one-click services.

Safety: Recipe creates are intentional — no confirm gate. No dry-run / preview. Partial failure does not auto-rollback; created UUIDs are returned in error.data. Connection strings are masked unless reveal: true. recommend is read-only / advisory.

recipe({ action: "create-git-app", server_uuid, git_repository, git_branch, repo_path: "/path/to/repo" })
recipe({ action: "create-app-db", server_uuid, app_name, db_name, db_engine: "postgresql" })
recipe({ action: "create-one-click", server_uuid, type: "gitea" })
recipe({ action: "recommend", stack: "Next.js + Postgres" })

Also use service.list-types to discover valid one-click type IDs before create-one-click.

🌱 Resource environment variables (envs:*)

Manage Coolify runtime configuration on applications, services, and databases through envs:* actions on the existing domain tools — no separate env MCP tool.

Confirm gates: envs:delete and envs:bulk-update always require confirm: true on all three tools. On application only, envs:sync requires confirm: true when applying (dry_run: false, the default) or when prune: true. envs:promote requires confirm: true when applying (dry_run: false).

Reveal policy: Env values render as *** by default. Pass reveal: true only after the human explicitly asks for plaintext — the agent must not auto-set reveal: true.

envs:sync semantics (application only): Supply exactly one of env_file (local path) or env_content (inline .env text). dry_run: true returns a diff (added, updated, unchanged, removed, optional conflicts) with no API writes; default dry_run: false applies changes. Remote keys missing locally are never deleted unless prune: true (also requires confirm: true). When local and remote values differ, set conflict_policy to overwrite, keep_remote, or abort after asking the human — apply with conflicts and no policy returns COOLIFY_CONFIRM_REQUIRED.

envs:promote semantics (application only, same instance): Product docs may say env.promote; the implemented action is application.envs:promote. Compare env vars between source_uuid and target_uuid (two applications in one Coolify instance — no cross-instance fan-out). Default dry_run: true returns preview buckets (only_in_source, only_in_target, value_mismatches) plus structured promotion_suggestions with follow-up tool/action hints; values are masked unless reveal: true. Applying copies into the target requires confirm: true. Default conflict_policy is keep_remote — mismatched target keys are skipped unless the human opts into overwrite or abort.

application({ action: "envs:list", uuid: "<app-uuid>" })
application({ action: "envs:sync", uuid: "<app-uuid>", env_file: "./.env", dry_run: true })
application({ action: "envs:sync", uuid: "<app-uuid>", env_content: "API_KEY=EXAMPLE_VALUE\n", confirm: true, conflict_policy: "overwrite" })
application({ action: "envs:promote", source_uuid: "<source-app-uuid>", target_uuid: "<target-app-uuid>", dry_run: true })
application({ action: "envs:promote", source_uuid: "<source-app-uuid>", target_uuid: "<target-app-uuid>", dry_run: false, confirm: true, conflict_policy: "keep_remote" })

💾 Database backups (backup:*)

Configure, list, update, delete, and trigger backup schedules — and inspect execution history — on the existing database tool. No separate backup MCP tool.

Parent identity: All backup actions require the parent database via uuid or name. Schedule-scoped actions (backup:update, backup:delete, backup:now, backup:history) also require scheduled_backup_uuid.

Confirm gates: backup:delete requires confirm: true — otherwise COOLIFY_CONFIRM_REQUIRED. delete_s3 defaults false (config-only delete). When delete_s3: true, deletion still requires confirm: true — purging S3 artifacts is treated as destructive.

Frequency (Pitfall 1): backup:create accepts OpenAPI named presets (every_minute, hourly, daily, weekly, monthly, yearly) or a cron expression. backup:update accepts presets only — passing cron on update returns COOLIFY_VALIDATION_ERROR.

backup:now semantics: Maps to Coolify PATCH with { backup_now: true } on the schedule — no separate trigger endpoint. Requires scheduled_backup_uuid.

Reveal policy: S3-related credentials in backup config responses are masked as *** by default. Pass reveal: true only after the human explicitly asks for plaintext — the agent must not auto-set reveal: true.

Out of scope (v2.x+): Backup execution delete, restore/import from backup, and S3 storage destination CRUD are not available in this release.

database({ action: "backup:list", uuid: "<db-uuid>" })
database({ action: "backup:create", uuid: "<db-uuid>", frequency: "daily", save_s3: false })
database({ action: "backup:now", uuid: "<db-uuid>", scheduled_backup_uuid: "<schedule-uuid>" })
database({ action: "backup:delete", uuid: "<db-uuid>", scheduled_backup_uuid: "<schedule-uuid>", confirm: true })

🔑 private_key — SSH key CRUD

Manage Coolify private keys with PEM content masked by default.

🖧 server — server CRUD & validation

📁 project — project CRUD

🌍 environment — environment CRUD

📚 docs — offline guides

🚨 emergency — high-impact ops (gated)

Reach for these only when you mean it — every action below is behind a confirmation gate.

🗂️ instance — multi-instance registry

Manage named Coolify instances in ~/.coolify-mcp/instances.json. Per-call credential resolution — no cross-instance leakage.

instance({ action: "add", name: "prod", url: "https://coolify.example.com", token: "<token>" })
instance({ action: "list" })
instance({ action: "cloud-info" })

📜 manifest — local cache

Read/write/sync .coolify/manifest.json — a workspace cache, not source of truth. Remote wins on UUID conflict.

manifest({ action: "sync", dry_run: true })
manifest({ action: "diff" })
manifest({ action: "audit" })

🧭 setup — guided project wiring

wire never auto-pushes. The setup flow pauses cleanly when gh authentication is missing and resumes from completed steps.

🎨 Branding (serverInfo.icons)

The MCP server advertises icons in initialize via an embedded PNG data URI (primary) and jsDelivr CDN URLs for mcp-icon-192.png and favicon-32.png. Cursor may still show a letter fallback — see maintainer verify record. Not a Coolify API call.

🛡️ Safety model

Confirmation gate

Destructive emergency actions follow a strict two-step pattern:

1. Call with confirm omitted or false → you get back a would_affect preview and error code COOLIFY_CONFIRM_REQUIRED — nothing is mutated.
2. Call again with confirm: true → the action actually executes.

Regular app/service/database mutations (start, stop, deploy, …) are not behind this gate — they simply follow Coolify's own API semantics, since they're scoped to one resource rather than your whole fleet.

Environment variables: envs:delete and envs:bulk-update require confirm: true on application, service, and database. envs:sync apply (dry_run: false) and envs:sync with prune: true require confirm: true on application only. dry_run: true sync previews never mutate. envs:promote apply (dry_run: false) requires confirm: true on application; default conflict_policy is keep_remote.

Drift & heal (read-only audit, preview-first promote): manifest.audit is advisory-only — it never writes manifest or live state. application.envs:promote (product name env.promote) previews by default; values stay masked unless reveal: true. Both stay within one Coolify instance per call.

Deploy guard (advisory preflight, confirm-gated rollback): deployment.preflight is read-only and returns a risk_score with four named factors — no external DNS/HTTP probes. deployment.rollback requires confirm: true before mutating; git rollbacks PATCH the target commit then POST /deploy (MCP composite, not a Coolify rollback API).

Secret masking

• Keys matching password, token, secret, private, or env render as *** by default in tool output.
• Pass reveal: true only when you explicitly need plaintext — for example, to copy an env var into another system. Ask the human first before setting reveal: true on any envs:* call.
• Log line bodies are not masked. Treat raw logs like you would any other sensitive output: don't paste them into long-lived agent memory or public tickets.

⚠️ Structured errors & retries

Every API failure comes back as a parseable envelope your agent can reason about, instead of a raw stack trace:

{
  "code": "COOLIFY_401",
  "message": "Unauthorized — invalid or expired API token",
  "recoveryHints": [
    "Verify the token in Coolify UI → Keys & Tokens",
    "Ensure the token has the required team permissions"
  ],
  "httpStatus": 401
}

Transient failures (HTTP 429, 5xx, or network errors) retry automatically up to 3 times with exponential backoff (1s → 2s → 4s) before giving up and returning the error to your agent.

💬 Example agent workflows

"Is my Coolify reachable, and what do I have?"

system({ action: "verify" })
system({ action: "infrastructure_overview" })
resource({ action: "list" })

"Find the nginx app, deploy it, then show me the logs."

resource({ action: "find", query: "nginx" })
application({ action: "deploy", uuid: "<uuid>", wait: false })
deployment({ action: "watch", deployment_uuid: "<deployment_uuid>", timeout: 300 })
application({ action: "logs", uuid: "<uuid>" })

"Something feels wrong across the fleet."

diagnose({ action: "scan" })
diagnose({ action: "app", uuid: "<suspect>" })
diagnose({ action: "server", uuid: "<server>" })

"Emergency: stop everything, but let me see the blast radius first."

emergency({ action: "stop_all" })                 // preview — would_affect, no mutation
emergency({ action: "stop_all", confirm: true })  // execute

"Multi-instance: list registered instances and verify each."

instance({ action: "list" })
system({ action: "verify" })

✅ Status today

Package 1.1.4 ships 19 tools and six MCP prompts for Coolify API 4.1.x:

Capability discovery & build logs: system({ action: "version" }) returns coolifyVersion (replacing the legacy version field), mcpVersion, and a capabilities map of Coolify 4.1.2 feature flags. For app triage + bounded runtime tail in one call, use diagnose({ action: "logs", mode: "full", uuid: "..." }) — check capabilities.diagnose_logs. For Log Brain pattern triage, use diagnose({ action: "analyze", uuid: "..." }) — check capabilities.diagnose_analyze. For advisory stack picks from the live catalog, use recipe({ action: "recommend", stack: "..." }) — check capabilities.recipe_recommend. For instance health, dependency graph, impact, and janitor/cleanup, use intelligence({ action: "scorecard" | "graph" | "impact" | "janitor" | "cleanup", ... }) — check capabilities.intelligence_scorecard (and sibling intelligence_* keys); cleanup requires confirm: true. For manifest drift audit and cross-app env promote, use manifest({ action: "audit" }) and application({ action: "envs:promote", source_uuid, target_uuid, ... }) — check capabilities.manifest_audit and capabilities.envs_promote (MCP composites over existing reads/env CRUD, not Coolify-native REST endpoints). For deployment build logs, prefer deployment({ action: "logs", deployment_uuid: "..." }) (or application_uuid to resolve the newest deployment). The application.logs path with deployment_uuid still works for back-compat. For runtime log follow, use application({ action: "logs", uuid: "...", follow: true }) — bounded MCP polling until idle or timeout; check capabilities.application_logs_follow via system.version.

🔮 Coming soon

Future work stays bounded by verifiable upstream and repository constraints:

• Add service/database logs when compatible Coolify APIs are stable and available.
• Close tracked REST mappings in docs/COVERAGE.md where they add useful agent workflows.
• Revisit cross-instance fan-out only with explicit rate-limit and credential-isolation guarantees.

No release date or compatibility promise is attached to these boundaries. Use GitHub Issues for concrete requests.

🛠️ Local development

git clone https://github.com/clezcoding/awesome-coolify.git
cd awesome-coolify
pnpm install
pnpm run build    # tsup → dist/
pnpm test         # vitest
pnpm run dev      # watch mode

Logs go to stderr only — stdout is reserved exclusively for the MCP protocol.

The maintainer publish flow (build → pack --dry-run → publish) is documented in CONTRIBUTING.md.

🔗 Links