OpenGriffin

The Claude-native personal agent that learns while you sleep — and lives in your Telegram.

You wake up. Overnight, your agent has read yesterday back to itself, written down what it learned about you, caught its own failures, and proposed new skills — in a journal you can read over coffee. That nightly self-improvement loop, built natively on the Claude Agent SDK — real skills, MCP, hooks, resumable sessions, not a reimplementation — is what OpenGriffin is. It lives in your Telegram, runs entirely on your own machine, and works with any of 21 model providers on your own key.

---

How it's different

The personal-agent space is crowded — Nous Research's Hermes Agent, OpenClaw, and QwenPaw all do self-evolving, multi-messenger, BYO-key, and local-first. OpenGriffin doesn't claim to out-feature them. Two things are genuinely ours:

1. It's the Claude-native one. OpenGriffin runs on the Claude Agent SDK itself — your skills are real Claude Code skills under ~/.claude/skills/ (shared with Claude Code), MCP servers, hooks, and resumable sessions all work as-is. The others reimplement an agent loop; OpenGriffin is the Claude loop, wearing a Telegram face.
2. The journal is a readable, shareable artifact. The nightly loop doesn't just mutate hidden state — it writes you a page you can read, and griffin card makes it shareable. "Learns while you sleep" you can actually see.

	OpenGriffin	Hermes (Nous)	OpenClaw	QwenPaw
Built on the Claude Agent SDK (native skills / MCP / hooks / sessions)	✅	❌ own runtime	❌ own runtime	❌ Qwen / AgentScope
Readable nightly journal you can screenshot & share	✅ griffin card	❌	❌	❌
Telegram-native onboarding (one messenger, ~2 min)	✅	one of many	one of many	one of many
Self-evolving skills	✅	✅	◑	✅
Multi-messenger gateways	✅ 7	✅ 7	✅ 20+	✅ 5
BYO-key / multi-provider	✅ 21	✅ (incl. 200+ via OpenRouter)	✅	✅
Local-first / no backend	✅	✅	✅	✅
Acts proactively (cron / triggers)	✅	✅	✅ heartbeat	✅
License	Apache-2.0	MIT	MIT	OSS

Comparison reflects publicly documented features as of mid-2026. Hermes, OpenClaw, and QwenPaw are excellent projects with far larger communities — this table is about fit, not size.

---

Why OpenGriffin

• Self-evolving skill graph. The agent authors, edits, and retires its own skills at runtime. No redeploy. New capabilities ship inside a conversation.
• Daily journal at 4:30 AM. Every conversation, decision, and tool call is appended to a structured journal. A nightly self-improvement loop reads yesterday's entries, summarizes them, and proposes new skills — then griffin card renders the entry into a shareable image.
• 21 AI providers, BYO key. Claude Max OAuth, Anthropic, OpenAI, OpenRouter, Azure, Gemini, Mistral, Cohere, Groq, Together, Fireworks, DeepSeek, HuggingFace, Perplexity, xAI, Bedrock, Cerebras, NVIDIA NIM, Lambda Labs, Novita, Ollama. Switch with one env var.
• Persistent memory: MEMORY / USER / SOUL. Three flat markdown files load fresh into every session. MEMORY.md is the environment, USER.md is the profile, SOUL.md is the voice.
• Multi-surface. Telegram is the front door, but the same brain answers via CLI, webhooks, and voice notes.
• Skill auto-discovery. Drop a markdown file into ~/.claude/skills/, and the agent picks it up on next turn — no registry, no config.
• Voice round-trip. faster-whisper transcribes inbound voice notes; edge-tts speaks replies in the user's preferred neural voice.
• Browser automation via Playwright MCP. Screenshots, scraping, end-to-end clicks. The agent can drive a real Chromium.
• Kanban built in. Multiple agent workers claim, block, and complete tasks against a shared board. Long plans get parallelized.
• Checkpoints + rollback. Conversation snapshots survive crashes. The agent can replay or branch off any prior point.
• Approval inline buttons. Risky actions ask first via Telegram inline keyboards. Confirm or deny with a tap; nothing destructive runs unsupervised.

---

This is the beginning of the README. Read the rest at the original. Read the full README ↗