Apache 2.0 · BYO key · No backend
OpenGriffin is an open-source personal AI agent. It lives in your Telegram, remembers everything, schedules its own work, builds its own skills, and runs entirely on your machine.
30 features. 21 AI providers, BYO any key. Daily self-improvement loop. Free forever.
30 features · 33 MCP servers · 11 nightly auto-jobs
Every feature is local-first. None require an account or a subscription. Bring your own AI provider key for any of 21 supported backends.
Vivid → recent → fading → ancient. Time-aware decay. Every recalled fact links to its source session via a citation receipt.
Three flat markdown files load fresh into every session. Environment, profile, voice. Edit them by hand or let the agent.
Hard rules the agent must NEVER violate. Loaded at the top of every system prompt, override-proof.
4:30am every day: agent reviews yesterday, writes a journal entry, consolidates memory, suggests skills.
3am offline reflection on interesting moments. Counterfactual analysis. Distills lessons into MEMORY overnight.
Flags when stated preferences contradict recent behavior. "You said you hated meetings; today you scheduled five."
griffin install github://owner/repo. License-checked, signed, reputation-tracked. Replaces tool sprawl.
When a skill fails 3+ times in a week, the agent debugs it, proposes a patched SKILL.md, and asks before applying.
Recommends new skills based on co-usage, flags never-used ones, surfaces top-used to invest in.
Long-running background agents. Spawn a research worker. Queue tasks. They check in to Telegram for days.
Multiple personas debate in a group chat with shared memory. Convergence detection.
Fork a child agent from a parent. Inherits SOUL, skills, memory snapshot. Diverges over time.
Cron + webhook + poll → LLM predicate → skill action. "When Stripe revenue drops 10% week-over-week, draft a postmortem."
Agent commits to do X at a future time. You can /veto before; otherwise it runs.
Detects time-of-day query patterns and pre-computes likely-asked things 15 minutes early. "I noticed you check NVDA at 8:30am — here it is."
Every consequential action reviewed by a second agent that doesn't see your prompt. Catches goal misgeneralization.
Signed, scoped, expiring permissions for tools. The agent can't do something it doesn't hold a token for.
Pattern-match for prompt injection, dangerous shell, exfil URLs, hardcoded secrets, homograph attacks. Hardline blocklist for fork bombs / mkfs / dd-to-disk.
Secure Enclave (macOS) / TPM (Linux) signing of every consequential action. Tamper-evident audit log.
Hash-chained Merkle audit log. Prove any single action existed without revealing the others.
If you don't check in for N days, outbound actions lock. Recovery code via Telegram. Optional escalation to a trusted contact.
High-stakes actions need 2-of-3 independent agent personas to agree. Reduces single-point prompt-injection failure.
Risky tool calls ask before running. Telegram inline keyboard: Allow once / Session / Always / Deny. 60-second auto-deny.
Every Write/Edit snapshots the file first. /rollback restores in one command.
Telegram, Discord, Slack, Email, iMessage, Signal, Matrix. Same brain answers across all of them.
Link your accounts across Telegram + Discord + Slack + ... Single memory namespace per human.
Your agent can call another OpenGriffin user's agent. P2P. Pay-per-call via x402 wallet, or free.
Signed JSON-LD profile of your agent's track record. A2A-discoverable for trust.
Telegram voice → faster-whisper → Claude → Edge-TTS → reply voice. Local STT/TTS, no cloud.
Playwright MCP. Drives a real Chromium for screenshots, scraping, end-to-end clicks.
FAL.ai wrapper for FLUX. BYO FAL_KEY. (More providers landing soon.)
JSON-backed task board. Workers + pods + ambient triggers all read/write it.
Server-Sent Events stream of every agent thought + tool call. Datadog APM for cognition.
Re-run any past session with a different model or SOUL. Counterfactual diffing for regression hunts.
Auction classifier scores each prompt 0–3 and routes cheap. "hi" → Groq. "prove this" → Claude Opus.
Mines your past chats to build a writing-voice card. Agent drafts as you.
HMAC-validated POST endpoints. GitHub, Stripe, IFTTT — anything that webhooks reaches your agent.
No central account, no shared pool. Your keys, your bills, your data.
No account. No waitlist. No backend. Apache 2.0.
curl -fsSL https://raw.githubusercontent.com/greentarallc/opengriffin/main/scripts/install.sh | bashOr with pip: pip install opengriffin && opengriffin run
Open Telegram, message @BotFather, run /newbot. Save the token to .env.
Drop any of the 21 keys into .env. Or use Claude Max OAuth — no key needed.
Bot starts. Send it a message. Memory persists. Tomorrow at 4:30am the journal writes itself.
Built-in migration imports your memories, cron jobs, and recent sessions in one command.
griffin migrate from-hermes
griffin migrate from-openclawImports MEMORY/USER/SOUL.md, cron schedules, channel directory, recent message history, and any local scripts.
Yes — Apache 2.0, free forever, self-hosted. You bring your own AI provider key and pay that provider directly. There is no OpenGriffin subscription, ever, on the OSS path.
OpenGriffin is a strict superset on memory (Echo Memory + receipts + dream cycle), self-improvement (daily journal + skill graph + drift detection), security (capability tokens + critic + Merkle audit + dead-man's switch), autonomy (worker pool + agent pods + genealogy + time-locked actions), and infrastructure (no backend, 21 providers, 7 platforms). Hermes has 14 messaging gateways and Atropos RL training; we don't (yet).
OpenClaw is a thin CLI wrapper. OpenGriffin is a long-running agent with persistent memory, scheduled work, sub-agent spawning, multi-platform delivery, and a self-improvement loop. We ship a one-line migration tool to import OpenClaw state.
Magai is a hosted multi-model dashboard. OpenGriffin is autonomy + memory + skills you run on your own machine. Magai costs $19+/month and your data lives on their servers; OpenGriffin is free and your data never leaves your disk unless you explicitly send it somewhere.
No. Telegram is the easiest first surface, but the same brain runs on Discord, Slack, Email, iMessage, Signal, and Matrix. Cross-platform identity links them.
100% local: ~/.opengriffin/ for memory + sessions + kanban + journal, ~/.claude/skills/ for skills. Nothing is uploaded. Outbound HTTP only happens to providers you configured (Telegram, Anthropic API, etc.) and to peers you explicitly call (A2A).
Eventually, when demand justifies the engineering. The OSS Core will always be free. The hosted product would offer optional managed hosting + a web GUI; the OSS will always do everything you can do hosted.
Layered defense: pre-execution scanner (blocks rm -rf /, fork bombs, exfil), capability tokens (signed scoped permissions), critic agent (independent review), approval inline buttons for dangerous patterns, hardware-attested signing, dead-man's switch, file checkpoints with one-command rollback. You can also run with permission_mode=plan to prevent any tool execution.
30 features. 21 providers. 7 platforms. Free forever.