What's new
What each release lets you do, in plain words. The newest is first.
Releases: 1.4.0 · 1.3.0 · 1.2.1 · 1.2.0 · 1.1.0 · 1.0.3 · 1.0.2 · 1.0.1 · 1.0.0 · 0.13.0 · 0.12.0 · 0.11.0 · 0.10.0 · 0.9.0 · 0.8.1 · 0.8.0 · 0.7.0 · 0.5.1 · 0.5.0 · 0.4.3 · 0.4.2
Restart from chat, bridge to Solana
Restart or update your agent from chat or one command, move funds into Solana, and keep your model choice on every run.
- Restart from chat Ask for a restart or an update in chat; the agent finishes its work, restarts and reports back.
- Bridge into Solana Move SOL or USDC from another chain into Solana, with the arrival proven on chain.
- Your model everywhere The model you pick now runs scheduled jobs, goals and planning, not only chat.
- Memory that recalls better Recall searches what you actually said, and memory stops storing tool noise.
- Simpler settings Fifty fewer settings; one line holds your overrides and the doctor names what to change.
A full terminal, a full mailbox, tighter guards
Your agent works a real terminal and its own mailbox, moderates its rooms, and stays inside stronger guards without asking you more often.
- A real terminal Long commands run as background jobs you can type into, on this machine or a remote sandbox.
- Its own mailbox The agent reads, replies, forwards, files and deletes mail in its own inbox.
- Room moderation The agent can mute, ban and delete messages in rooms you allow, and remove its own posts.
- Fewer taps Trusted buys, cancels and sends to your own addresses go through without asking you.
- Stronger guards Every shell command passes one guard, and wallet limits never exceed what the signer allows.
Safer chat, steadier delivery, agent NFT journal
A fix release: messages arrive once and in full, chats and the console say when they cannot read, and money waits survive a cancel.
- Messages arrive once A half-sent message resumes instead of repeating, and quiet hours keep the full text and its files.
- Honest reads The console and chats say when they could not read something, instead of showing it as empty.
- Safer money waits Stopping a money move mid-wait still counts it and tells you it may have gone through.
- Edited messages stay edits Editing an old chat message never runs its command a second time.
- Agent NFT journal Your agent NFT keeps a signed history of its actions, and you can trust a collection from chat.
One install, more chat apps, money from chat
One install with optional packs, three more chat apps, and money moves you confirm with one tap from any chat.
- One-command installer One command installs POLYROB on Linux, macOS and WSL2 and runs the setup wizard.
- Packs Add or remove optional features as packs: tools, commands and skills in one install.
- More chat apps Talk to your agent in Feishu, Lark and DingTalk; approve with buttons and hear voice replies.
- Money from chat Send, swap and pay from chat with one-tap cards; the pause stops only the agent, caps bind everyone.
- Check any token Ask about any wallet or token from chat and get a sourced, plain answer.
Questions that wait for your answer
Autonomous work asks you real questions, uses focused toolsets and shows what it costs.
- Questions that stay open When a job needs a decision, it asks on every owner channel and waits for you.
- Answer in your own words Reply to a question in writing; the job retries with your answer.
- Focused toolsets Give each scheduled job or goal a named toolset such as money, research or social.
- Costs you can see Status shows calls, dollars per day and how long your budget will last.
- About 270 fixes Audits of the console, API, terminal and chat apps found about 270 defects; this release fixes them.
X on servers and fair scheduling
Move an X login to a headless server and keep your chat turns ahead of background work.
- X login on servers Log in to X on your desktop, then move that login to a headless server.
- Reply on X The agent can reply under X posts, after you approve each reply.
- Goal step budgets Give each goal its own step budget, from 6 to 60 steps.
- Your turn comes first While you chat, scheduled jobs and updates wait until your turn ends.
- Safer updates Updates no longer touch other git folders, and a full context no longer ends a run.
Wallet limits from chat
Change wallet limits from chat with immediate effect, and run the browser as an isolated service.
- Wallet limits from chat Set the per-transaction limit from chat; your daily cap stays the hard maximum.
- Limits apply at once A new wallet limit takes effect immediately, with no restart.
- Isolated browser Run the browser as a sandboxed service with its own user and restricted network.
- Deploy on Fly.io A ready container setup runs the agent on Fly.io with no public ports.
- Cleaner chat replies No second recap after an answer, and replies from Gemini models no longer go missing.
An agent that sets itself up
A fresh agent checks its own setup, and status shows who it is and what it holds.
- Self-setup A fresh agent checks its setup, fixes what it can and sends you one list of the rest.
- X poll results The agent can read the votes on polls it posted.
- Encrypted X Chat It can read encrypted X Chat history as well as normal direct messages.
- Identity in status Status shows the agent's email address and whether its X account is ready.
- Liquidity positions Status, the wallet overview, Telegram and the console list liquidity positions and fees.
First stable release
Terminal, console, chat apps, autonomy, tools and an optional wallet under one clear owner model.
- Terminal agent Chat or run one-shot tasks with files and images; pause, steer or stop a busy session.
- New web console One console with Chat, Work, Money, Inbox and Agent views and a command palette.
- Telegram groups Groups get roles, limited toolsets and reply threads; only the owner administers them.
- Owner stays in control A stop survives restarts, and only the owner can resume autonomy.
- Optional wallet Guarded swaps, bridges, token deploys, NFTs, on-chain identity and x402 — off until you allow them.
One stop button and live apps
Stop all autonomous work at once, keep built apps online and get paid without a server.
- One stop for everything /pause from any channel stops all autonomous work in seconds, and the pause survives a restart.
- Scoped pauses Pause only trading, background work or messages, optionally for a set time.
- Apps that stay online A built app runs in its own container at a public address after you approve it once.
- Payments without a server Invoices use the agent wallet, and USDC payments on Base are detected automatically.
- Honest status Every status view starts with a health summary: credit, open questions, blocked goals, pause.
Several agents on one machine
Run isolated agent profiles side by side and share them as packages.
- Named profiles Create isolated agents, each with its own keys, persona, skills, memory and goals.
- Shareable agents Export or install a profile from git; keys and memories never ship with it.
- Neutral default persona A fresh install starts as a neutral POLYROB agent.
- x402 discovery Check what a paid endpoint costs before you pay, with no wallet needed.
- Unattended trading Scheduled goals can trade within set limits, and only if a daily cap is set.
Its own email, X account and web pages
The agent can publish pages, own an inbox and an X account, and setup is simpler.
- Publish to a URL The agent puts a built page at a real web address after you approve the first publish.
- Its own email One AgentMail key gives the agent its own inbox, with no mail server setup.
- Its own X account An opt-in browser tool registers and posts on X; you approve each post.
- Easier provider setup A guided prompt connects model providers; Ollama Cloud, z.ai and Cerebras are built in.
- Better first run A plain pip install works, database updates apply automatically and errors show one clear fix.
Any model provider, token tools
Add any compatible model provider without code, and give the agent guarded token access.
- Bring any provider Add Ollama, LM Studio, vLLM, Groq or any compatible endpoint in one YAML file.
- Token lookup A read-only tool shows token prices, liquidity, safety checks and holdings.
- Guarded transfers Each transfer is simulated and capped; larger amounts wait for your approval.
- Fewer wasted calls When credit runs out, the agent stops at once instead of making more paid calls.
- Security gates repaired Several approval and safety checks that never matched now work.
Spend budgets and MCP server
Cap what a run may spend, connect MCP clients and chain goals in order.
- Per-run budget Set a dollar limit; a run stops honestly when it reaches it.
- MCP server Claude Desktop, Cursor and other MCP clients can read usage, goals and pending approvals.
- Goals that wait A goal can start only after other goals finish, and retries after an outage.
- Remote code execution Run agent code on a remote host over SSH.
- Autonomy off by default A new install does no background work until you switch autonomy on.
Tools on demand, files in chat
The agent loads the tools it needs, and finished work arrives as files in chat.
- Tools on demand Each session sees every tool and can load one when it needs it.
- Files with results Finished goals and scheduled jobs attach their output files in Telegram.
- Search from chat Search the knowledge base and list recent files from Telegram.
- Console links Completion messages link straight to the session in the web console.
- One-time avatar Roll a face and voice for your agent, then keep it for good.
Live progress and built-in payments
See what the agent does as it happens, and invoice and get paid in USDC.
- Live progress Telegram, the terminal and the console show the current step, tool, cost and waits.
- Autonomous mode One setting makes a single-owner agent capable; spending and secrets stay locked.
- Invoices and payments Send QR invoices, detect USDC payments and run prepaid subscriptions.
- One-command wallet Create or import a standard 24-word wallet that you can back up.
- Settings you can find Search, explain and change any setting from the terminal or the console.
New models, cleaner terminal
New Claude, Grok and GLM models, a fixed terminal display and email threads that stay together.
- New models The latest Claude models plus new Grok and GLM options, with verified prices.
- Clean terminal Background logs no longer break the chat screen.
- Real email threads Each email reply goes back to the conversation that sent it.
- Conversations resume A reply to a closed session opens a new one with the full history.
- Same controls everywhere The console and the terminal gain the same settings and review pages.
Wallet and update fixes
The wallet spends from the address it asks you to fund, and updates work on pinned installs.
- Fund equals spend The agent wallet now spends from the same address it tells you to fund.
- Updates on pinned installs Update now works on an install that follows a release tag.
- Small fixes Minor reliability fixes across the wallet and the updater.
Money loop and full-control console
The agent can invoice and get paid, you choose how much of the machine it may use, and the console sees everything.
- Invoices and payments The agent can send invoices, get paid in USDC and show what it earned.
- Machine access levels Choose how much of the computer the agent may use, from none to approved self-maintenance.
- Full-control console One console shows every session, a live activity feed and lets you step in.
- Restart-safe autonomy Autonomous work continues after a restart, and wake-ups happen only when something changed.
- Settings health check A doctor command lists every setting and what it does.
Honest goals and owner questions
Goals can now honestly fail, and the agent asks you when it is stuck.
- Goals are checked A finished goal is verified; if the work is not done, the goal fails honestly.
- The agent asks you When a goal blocks, the agent asks you once instead of stalling in silence.
- Approve from Telegram List, approve and reject pending actions from Telegram or the terminal.
- Message tool The agent can send messages, only to an allowlist you set.
First public release
A self-hosted autonomous agent that pursues goals, learns from experience and runs on your own machine.
- Goal-driven agent Give it a goal in plain language; it plans, browses, reads, writes and runs code.
- Memory that lasts Cross-session recall with hierarchical memory that forgets by importance.
- Durable autonomy A goal board that survives restarts, schedules, self-wake and self-written skills.
- Many interfaces Terminal, web dashboard, REST API, A2A, MCP and chat apps.
- Safe by default Untrusted input is marked, sub-agents get fewer tools and strangers get limited access.
Older releases and every fix: the full CHANGELOG.md on GitHub.