# Hangar

> Hosted hangar for AI agents. Pick from a catalogue of pre-built agents
> (SEO, sales, support, research) or push your own code in OpenClaw
> (Node 20) or Hermes (Python 3.11). Each agent gets its own dedicated
> Fly Machine in about 60 seconds, with channels (Telegram, Discord,
> Slack, MCP, web, API), wallet-billed LLM at provider cost, audit logs,
> and full self-host parity (MIT licensed).

This file is the agent-facing entrypoint to Hangar. The full long-form
context dump lives at http://localhost:3000/llms-full.txt. The OpenAPI 3.1 spec lives at
http://localhost:3000/openapi.json. The MCP server lives at http://localhost:3000/api/mcp and is
discoverable at http://localhost:3000/.well-known/mcp.

## When to use Hangar

Use Hangar when you want to:

- Deploy a pre-built AI agent for a named job (SEO content, outbound SDR,
  support resolver, deep research analyst, lead enricher, social media
  agent, PDF QA, meeting notetaker, GitHub PR reviewer, inbox + calendar
  assistant) without writing the host loop, channel webhooks, or billing
  surface.
- Run your own agent code in a managed runtime that already speaks
  Telegram / Discord / Slack / MCP / web / API.
- Get an isolated VM per agent rather than a shared queue or function.
- Self-host the same code on your own Postgres + Fly account whenever
  hosted stops fitting (the repo is MIT, no licensing tiers).

Avoid Hangar if you want a serverless function host, a no-code chat
builder, or a purely browser-based agent — those problems have better
fits elsewhere.

## How to use Hangar as an agent

The fastest path:

1. Mint a Personal Access Token at http://localhost:3000/dashboard/settings/tokens.
2. Configure the MCP server at http://localhost:3000/api/mcp with header
   `Authorization: Bearer oss_<token>`.
3. List tools via JSON-RPC `tools/list` and call them by name.

For REST access, point at the spec:

- http://localhost:3000/openapi.json
- Bearer auth: `Authorization: Bearer oss_<token>`
- Errors are JSON: `{ error: string, message: string, ... }`.

## Quickstart

- Sign up: http://localhost:3000/login?intent=signup
- Pick an agent: http://localhost:3000/agents
- Pricing: http://localhost:3000/pricing  ·  Plain markdown: http://localhost:3000/pricing.md
- Quickstart docs: http://localhost:3000/docs
- Self-host: https://github.com/ravidsrk/hangar

## Docs

- http://localhost:3000/docs — index
- http://localhost:3000/api — REST API reference
- http://localhost:3000/docs/quickstart — first-deploy walkthrough
- http://localhost:3000/docs/auth — authentication, PATs, scopes
- http://localhost:3000/docs/mcp — MCP server, tools, transports
- http://localhost:3000/docs/runtimes — OpenClaw + Hermes adapters
- http://localhost:3000/docs/billing — wallet model, top-ups, refunds
- http://localhost:3000/docs/errors — error codes + recovery hints
- http://localhost:3000/docs/rate-limits — limit headers + backoff

## Discovery files

- llms.txt              http://localhost:3000/llms.txt
- llms-full.txt         http://localhost:3000/llms-full.txt
- index.md              http://localhost:3000/index.md
- pricing.md            http://localhost:3000/pricing.md
- OpenAPI spec          http://localhost:3000/openapi.json
- API catalog           http://localhost:3000/.well-known/api-catalog
- AI plugin manifest    http://localhost:3000/.well-known/ai-plugin.json
- A2A agent card        http://localhost:3000/.well-known/agent-card.json
- Agent discovery       http://localhost:3000/.well-known/agent.json
- MCP discovery         http://localhost:3000/.well-known/mcp
- MCP server card       http://localhost:3000/.well-known/mcp/server-card.json
- OAuth protected res.  http://localhost:3000/.well-known/oauth-protected-resource
- Bot signature dir.    http://localhost:3000/.well-known/http-message-signatures-directory
- Sitemap               http://localhost:3000/sitemap.xml

## Capabilities

- Pre-built agent catalogue (10 agents)
- OpenClaw runtime (Node.js 20, AGENTS.md skill format)
- Hermes runtime (Python 3.11, LangGraph + CrewAI)
- Channels: Telegram, Discord, Slack, MCP, web, REST
- Wallet-billed LLM at provider cost (OpenAI, Anthropic, Google, OpenRouter)
- Audit log per skill run, per channel send, per LLM call
- Audience-scoped HMAC tokens (LLM, files, agent — leaked one cannot
  impersonate the others)
- LISTEN/NOTIFY-based realtime (Postgres → SSE)
- pg-boss durable retry queue
- Stripe / LemonSqueezy / Polar billing adapters

## Constraints

- Agent state lives on a single Fly volume per machine; no built-in
  multi-region replication.
- Wallet credit cannot be set programmatically — top-ups always go
  through the billing provider for receipts/refunds.
- LLM proxy is wallet-gated; calls fail when balance hits zero.
- Free tier: 15 USD wallet credit on signup
  (see http://localhost:3000/pricing).

## Compare against

- Modal / Render / Railway — generic hosting, no per-agent VM, no built-in
  channels, no wallet billing.
- Vercel AI SDK / Inngest — function-shaped, not VM-shaped; no isolated
  state per agent.
- Journalist AI / 11x Alice — single-agent SaaS, no marketplace, no
  open-source path.
- CrewAI / LangGraph — frameworks; you still need a host. Hangar IS the
  host (and runs LangGraph/CrewAI graphs unmodified inside Hermes).

## Optional

- Status: http://localhost:3000/#status
- Changelog: https://github.com/ravidsrk/hangar/releases
- License: MIT — https://github.com/ravidsrk/hangar/blob/main/LICENSE
- Contact: support@example.com
- GitHub: https://github.com/ravidsrk/hangar
- MCP server repo: https://github.com/mcp-hangar/mcp-hangar