polyrob
Open source · MIT · runs on your machine
It browses, builds and follows up — in your chats, on your machine, with any model.
Ask in plain words. Watch every step, its time and its cost.
Telegram, WhatsApp, Discord, Slack, Signal, email, terminal and web. One agent, one memory, wherever you talk.
A goal board that survives restarts. With autonomy on, it schedules runs and follows up.
It edits code in a real shell, tests it in a sandbox and puts the app live behind a public URL you approve.
A real browser that logs in, clicks and fills forms — plus structured data from popular sites.
Its own wallet on EVM and Solana. Pays and charges over x402, capped per day. Opt-in, unaudited.
REST, A2A, MCP, /v1. Per-user isolation, metering, credits. One MIT install, no seats.
Local memory across sessions. With autonomy on, it drafts skills from what worked; you review them first.
It hands subtasks to parallel sub-agents with a narrower toolset.
Native apps for desktop and phone, and a hosted cloud for people who do not want to run a server. Self-hosting stays free.
Opt-in. Unaudited. Start on a testnet. Every on-chain write goes through the same checks.
Same memory, same goals, everywhere. Send it photos and files; it sends files back.
$ polyrob chat # talk to it $ polyrob run "…" # one task, then exit $ polyrob gateway # every channel, one process › /inbox # what needs you › /goals # what it is working on › /model # pick the model › /pause # stop all autonomous work
Use any MCP server, or serve its tools to others.
Optional Modal, Daytona and Vercel packs.
Read, reply, forward, file. Every recipient gated.
Posts and DMs. Posts wait for your tap.
Approvals on every channel. They fail closed.
Plus a knowledge base from your folders.
Send a voice note; it reads it.
Own keys, memory and goals per profile.
/v1 and MCP serverSKILL.md) · MCP servers · custom model providers · packsYes. The engine is MIT and self-hosting is free, with nothing behind a paywall. You pay only your model provider — or nothing with a local model. A paid cloud is coming for people who do not want to run it themselves.
OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter and NVIDIA NIM are built in. Any OpenAI- or Anthropic-compatible endpoint (Ollama, LM Studio, vLLM, a company gateway) goes in one providers.yaml.
Only if you ask. A fresh install answers you and does the task you give it. Switch autonomy on and it keeps a goal board, runs schedules and follows up.
The wallet is off by default and unaudited. When on, every write is simulated and checked against what the agent declared, then a per-transaction limit, a daily cap and your approval queue apply. Start on a testnet with a small cap.
Send /pause from any channel, the CLI or the console. It stops all autonomous work, or one scope such as the wallet, and survives a restart. /resume starts it again.
It stays on your machine. Calls go out to your model provider and the tools you enable. The installer sends one anonymous count after a fresh install; --no-telemetry or DO_NOT_TRACK=1 turns it off.
It edits code and runs tests, but it is a general agent: it browses, messages, schedules, remembers and can hold a wallet — as a long-lived service, not only in one terminal.
Yes — see the Hermes and OpenClaw migration guides and the dated comparison.
Yes. Run it as a service with the console, the REST API or the OpenAI-compatible API. Per-user isolation, metering and credits are built in. Isolation is a software boundary; run separate instances for hostile tenants. See deployment postures.
Yes. Named profiles give each agent its own keys, persona, skills, memory and goals on one machine.
One command. It finds Python 3.11+ and runs a setup wizard.
Paste any provider key, or point it at a local model.
polyrob chat to talk, polyrob run "…" for one task.
Prefer pip? pipx install "polyrob[all]" — see the install guide.