POLYROB runs on your machine — self-hosted, no hosted version. Targets Python 3.11+. Bring a key for any LLM provider you have.
Linux, macOS and WSL2. It finds Python 3.11+, clones the source into
~/.polyrob/src, builds a virtualenv, puts a polyrob command on your
PATH, and runs the setup wizard — the wizard works inside the pipe, so you are asked
for your provider key right there. It installs the current public release and adapts to it, so
what it prints at the end is what that version actually has.
Rather read it first? That is the better habit:
Useful flags: --extras server,browser · --browser /
--no-browser · --no-setup · --no-prompt (fully scripted) ·
--branch / --commit (pin a version) · --uninstall.
Or a leaner install — see the extras table below.
Prefer the newest source? Install straight from GitHub:
For a normal virtualenv install, use python -m playwright install chromium inside that environment.
Provider keys, default model, toolset, persona, owner pairing and the
autonomy dial. Writes ~/.polyrob/.env.
Verifies Python, config, keys (redacted), provider/model, workspace, database, memory and ports — and names the one command for anything missing.
A terminal REPL with the full agent. One-shot instead? polyrob run "your task".
Keeps the extras you installed with, snapshots before it changes anything,
and waits for the agent to be idle. polyrob update --check only looks, and
polyrob update --rollback restores the last snapshot.
Idempotent: it fast-forwards the checkout, reinstalls, and leaves your
config and memory alone. --branch or --commit pins a version.
Removes the command, the PATH entry, the virtualenv and the
source. Your ~/.polyrob is kept — it holds your keys and the agent wallet seed,
and your memory lives in a .polyrob beside each project you ran in.
Two switches, deliberately. The first turns on the tools you drive on a single-user machine — coding, git, knowledge base, cross-session memory, project context, messaging:
The second turns on the loops that act without you — self-wake, the goal board and planner, the curator, background review, and the agent writing its own skills and self-model. A fresh install is interactive-only until you ask for this:
(Set automatically when you choose an autonomous posture or mode, so a deliberate autonomy choice is never left inert.)
| Extra | What it adds |
|---|---|
(none) | Core agent, keyword memory, CLI, the OpenAI SDK (serves OpenAI, OpenRouter, NVIDIA, DeepSeek and every OpenAI-compatible endpoint). On your own machine an optional extra installs itself on first use. |
gemini | The Google Gemini SDK |
anthropic | The Anthropic SDK (also the z.ai / Anthropic-transport seats) |
server | FastAPI REST API + POLYROB Console, upload MIME sniffing |
browser | Playwright browser automation |
docs | PDF and DOCX reading |
media | Invoice-card QR, the avatar's Pillow renderer, run-history GIFs |
anysite | The AnySite scraping CLI |
memory-vector | Semantic vector recall for H-MEM (local local_vector backend: sqlite-vec + numpy + the embedder) |
crypto | Agent wallet, x402 pay-per-request, EVM on-chain verbs (swap / bridge / deploy / launchpad / dapp), Hyperliquid / Polymarket |
solana | Solana rail — Jupiter swaps, bridging and SPL token deploy (its own extra: a second curve and a second transaction model) |
telegram | Telegram chat surface |
twitter | Twitter/X outreach tool |
voice | Local voice transcription |
all | Everything above |
Chat, Inbox, Work, Money and Agent at http://localhost:5050. The default loopback posture needs no login; public deployments use an authenticated posture.
Local REST API (+ A2A + OpenAI-compatible /v1) at http://localhost:9000.
Building an agent product on top? The multi-tenant, metering and on-chain-payment core is built in — see Deployment and the full capabilities.
