Sky agent

Deploy your own AI agent to the cloud in one click.

Deploy

1. Fork this repo.
2. On Railway: New Project → Deploy from GitHub repo → pick your fork.
3. Add LLM_API_KEY - this can be an API key from any provider, including Anthropic, OpenAI, OpenRouter, etc.
4. Add TELEGRAM_BOT_TOKEN (see below)
5. Deploy.

Railway auto-redeploys every time you push to your fork, so editing your agent = editing files on GitHub.

Get a Telegram bot token

1. Open Telegram and message @BotFather.
2. Send /newbot, pick a name and a username.
3. Copy the token it gives you — that's your TELEGRAM_BOT_TOKEN.

Telegram enables itself when TELEGRAM_BOT_TOKEN exists. In chat: /reset wipes the bot's memory of that conversation.

Make it yours

Want to change...	Edit...
API endpoints	channels/http_api.py
Personality, model, context length	agent.yaml
The agent's abilities (tools)	agent/tools.py
Channels (add Discord, Slack, ...)	drop a new file in channels/

Give your agent files

In the GitHub web UI open knowledge/, Add file → Upload files, commit. After the auto-redeploy, ask your agent "what files do you have?"

Plain-text formats work (md, txt, csv, code). PDFs and images are detected but not supported for now.

How it works

Telegram / HTTP ──▶ respond(chat_id, text, config) ──▶ Claude API
                          │                              │
                     SQLite memory ◀──── tool use loop ──┘

Conversation memory is SQLite on disk, mount a volume (DB_PATH) to keep it across redeploys; on Railway, add a volume mounted at /data and set DB_PATH=/data/agent.db.

Self-optimizing benchmark

This repo also ships a benchmark harness where a meta-agent rewrites its own multi-agent pipeline — prompts, tools, topology — to hill-climb on Harbor tasks. See benchmark/ to run an experiment.

License

MIT — do whatever you want with it. Have fun.