Skip to content

OpenSuperBot v0.5.0: each teammate on its own AI

Choose a tag to compare

@wolfqing wolfqing released this 08 Nov 09:14
· 55 commits to main since this release

Each teammate can run on its own AI: Claude, OpenAI, xAI, Gemini, MiniMax and other OpenAI-compatible services, each connection with its own key. Also a Chinese UI, local computers and start at sign-in on Windows, CI on Linux, macOS and Windows, a first slice of browser workflow recording, and several security fixes.

OpenGrokBot is now OpenSuperBot, a name that no longer borrows xAI's "Grok". Existing installs upgrade in place: see Coming from OpenGrokBot.

Security

  • A local computer's screen stream (/screen/:botId/ws) was served without the gateway's token and origin checks; it is now guarded like the noVNC path.
  • Sandboxed commands can no longer read the operator's home directory. Both sandboxes hide it entirely and put back only runtime managers (~/.nvm, ~/.volta, ~/.pyenv, ~/.cargo/bin, …) read-only; before, Seatbelt denied a short list of key folders and bubblewrap denied none of them. The self-check now also fails if a command can list the home directory, and the switch refuses to turn on when no self-check is wired.
  • write_file on a local computer opens the final path component without following symlinks.
  • On Windows the gateway token file is limited to your account: its access list is reset (which also clears entries written on the file itself, such as the creator's default SYSTEM and Administrators grants), inherited access is removed, and only your user (by SID) is granted. Before, 0600 did nothing there and the file kept whatever its folder allowed. Start-up warns if it cannot restrict the file, and an existing token is tightened on the next start.
  • A local computer's file tools refuse Windows device names (CON, NUL, COM1, … — reading CON would wait for input forever) and colons, which would write into an alternate data stream.
  • The gateway runs on hono 4.13.16 and @hono/node-server 2.1.6, which fix several advisories. The one that reached the gateway: a client that dropped its connection while the gateway was still setting up a WebSocket (a screen viewer while the teammate's browser starts, say) raised an unhandled socket error and stopped the process (GHSA-qv5m-9pm6-hqp3). The server also no longer turns a backslash in the path into /, which could carry a request past a reverse proxy's path rules (GHSA-jpvv-r467-5p4w). The rest do not apply here: the origin check reads the Host header rather than the request URL (GHSA-7v5p-43qr-wch5), the token check is a single * middleware, and the gateway uses neither mount() nor the form validator. ws 8.22.1, @types/node 26.6.6 and @vitejs/plugin-react 6.2.1 came in the same update.

Added

  • Each teammate can run on its own AI. AI models (the AI button next to + in the sidebar) lists your connections — a service, its API key and a default model — and adds, tests, edits or deletes them; the top of a teammate's thread shows which connection and model it uses, and a click switches it from its next turn. The connection set up first (the settings and env vars from before) stays the default for every teammate that has not picked one. Schema v7 adds a connections table and two columns on bots; each extra connection's key lives in <data>/llm-keys/<id>, never in the database. Only your account can read it (mode 0600 in a 0700 folder; on Windows, an ACL for your account alone), and the API only ever returns it masked. Deleting a connection moves its teammates back to the default.
  • Claude through its own Messages API (the official @anthropic-ai/sdk), not the OpenAI-compatible shim: tool calls, screenshots, and Claude's thinking blocks handed back untouched within a turn. Requests use automatic prompt caching, and on Anthropic's own API the server-side refusal fallback for the models that have it; a decline that remains becomes a readable error instead of an empty reply. A custom connection can also speak this API, for services that offer an Anthropic-compatible endpoint.
  • The OpenAI Responses API as a third way to talk to a model, with the reasoning items handed back within a turn (store: false, encrypted reasoning).
  • More presets, checked against each provider's docs on 2026-11-07: Claude (claude-opus-5-5, effort medium), Gemini through AI Studio's OpenAI-compatible endpoint (gemini-3.5-flash-lite), MiniMax for the global and China sites (MiniMax-M3.1-Flash, reasoning effort low instead of its default max), and MiniMax M Plan, the one subscription plan whose terms let any compatible tool connect: the form shows its terms and wants a tick before it saves one. The GLM, Kimi and Alibaba coding plans are left out on purpose; their terms forbid use from your own apps, bots or automation. OpenAI and xAI now talk over the Responses API: every GPT-6 model can call tools there (on Chat Completions most cannot), OpenAI's preset asks for reasoning effort low, and xAI marks Chat Completions deprecated. AI models lists the services and which subscriptions other apps may use.
  • A Claude Pro/Max sign-in token (sk-ant-oat…) is refused wherever a key is saved or tested, with a note to create an API key in the Claude Console instead: Anthropic does not allow other apps to use subscription sign-ins.
  • GET/POST /api/connections, PATCH/DELETE /api/connections/:id, POST /api/connections/test, GET /api/connections/:id/models and PUT /api/bots/:id/llm. The default connection keeps the settings rules: fields pinned by env vars stay locked. Changing a connection's address without a new key drops the stored key, a test against a new address never sends the old key, and moving a connection to another service resets the models teammates picked on it. A Claude connection without a key refuses to run instead of letting the SDK pick up Anthropic credentials or ANTHROPIC_CUSTOM_HEADERS from the machine; a reply cut off at the output limit is reported instead of acted on.
  • The gateway keeps settings of its own (schema v6 adds a settings table): provider, API base, model, vision, language, computers, browser mirror and update checks. An env var still wins over a saved value. The model API key can live in <data>/llm-key instead of OPENSUPERBOT_API_KEY. Like the gateway token next to it, only your account can read it (mode 0600; on Windows, an ACL for your account alone), and it is out of reach of sandboxed commands and containers. The AI models dialog writes the model settings; the rest have no UI yet.
  • Model settings apply without a restart: the API base, model, key and vision are read again before each model call, so a change takes effect on the next call while a call already in flight finishes as it started. The choice of computers and the browser mirror still apply at startup.
  • Settings API for the coming wizard and settings page. GET /api/settings returns each value and where it comes from (env, saved, default); the key is only ever returned masked (sk-…3456), with its source (env, file, none); needsRestart lists the saved changes that wait for a restart. PUT /api/settings saves a batch, key included, or nothing at all: values pinned by env vars are refused as locked. POST /api/settings/test tries a provider before you save, first GET /models and then a 16-token chat, and sorts failures into codes (bad_key, quota, rate_limited, model_not_found, unreachable, timeout, provider_error). The saved key is only sent when the API base is unchanged, so one provider's key never goes to another. GET /api/settings/models lists the saved provider's models.
  • Provider presets (GET /api/settings/providers) for DeepSeek, Kimi, Qwen (Model Studio, Beijing), GLM, xAI, OpenAI, OpenRouter, a local Ollama and any OpenAI-compatible endpoint: address, where to get a key, a default model that calls tools, whether it takes images. They were checked against each provider's docs on 2026-11-03. A preset can add request fields the provider needs: DeepSeek and Qwen get thinking turned off; OpenAI's GPT-6 models only call tools on Chat Completions with reasoning_effort: "none". The connection test sends the same fields, and treats a 400 that names the key (xAI's answer to a wrong key) as bad_key.
  • GET /api/setup tells the coming first-run wizard whether it is needed and what the computers have to work with. It is not needed once there is a key, a provider that needs none, the demo model, or any model env var, so existing setups never see it. It reports Docker, the sandbox, the browser found or missing, and how big the download is (about 190 MB, 370 MB on disk). In a mainland China time zone it suggests the npmmirror.com mirror. POST /api/setup/browser downloads the browser ahead of time, and /ws pushes setup.browser progress (which item, percent, size) as it goes.
  • The first local computer and the wizard share one browser download instead of possibly running two into the same folder. The browser mirror is read when the download starts, so changing it no longer needs a restart.
  • With computers on auto, a gateway whose Docker is running but whose bot image is not built now uses local computers, instead of handing out Docker computers that fail on first use. An image built before the rename (opengrokbot/bot:dev) counts as built. The log says why and gives the docker build command. GET /api/setup adds docker.imageReady and the reason the computers were picked (why: chosen, docker_running, docker_not_running, docker_image_missing), and a choice made on the settings page is logged as such rather than as OPENSUPERBOT_COMPUTERS. Settings ask for a restart only when it would change the computers: switching auto to local on a gateway that already runs local computers does not.
  • The UI speaks Chinese. Every label, button and hint in the web UI comes from a dictionary (web/src/i18n/en.ts, with zh-CN.ts beside it); the language is the one saved in settings, otherwise the browser's, otherwise English. Dates and numbers follow it too. A missing Chinese entry fails the build, and a test checks that each translation fills in the same values as the English. Word choices are in web/src/i18n/GLOSSARY.md. What teammates say stays in their own words.
  • Reasons the gateway sends back now carry a code the UI translates, next to the same English as before: why a computer did not start (Docker not running, bot image not built, browser missing or failed to download), why commands cannot be turned on, why a release cannot become a rule, a taken teammate name. The lines the gateway posts for a teammate (stopped, hit an error, over budget, paused at a limit) keep their English text for the model and add a notice payload; schedules are described in the UI's language from the cron expression.
  • Teammates write in the operator's interface language when one is saved in settings, and otherwise in the language the operator writes in.
  • First teach-mode slice: record semantic browser actions on the teammate's existing page, review the draft before saving it to SQLite, redact sensitive input, and reattach across local-computer window relaunches. API-only; replay and routines remain out of scope.
  • Local computers on Windows, with the browser and the folder: the gateway finds Chrome, Edge, Chromium or Brave where Windows installs them (for all users or just you) and otherwise downloads Chrome for Testing. Commands stay off, because Windows has no sandbox as strong as Seatbelt or bubblewrap; a teammate that needs a shell there gets a Docker computer.
  • Start at sign-in on Windows: deploy/windows/install-task.ps1 registers a Task Scheduler task that runs the gateway as you, in your session (so a sign-in window can appear), and restarts it if it stops.
  • CI on Linux, macOS and Windows. Typecheck, tests and build run on all three, and a new job runs the local-computer smoke test on each against the runner's real Chrome — on Linux with bubblewrap and a virtual display, on macOS with Seatbelt, on Windows with the extra path checks. The Windows job also installs and removes the start-up task.
  • deploy/linux/bwrap.apparmor: the AppArmor profile bubblewrap needs on Ubuntu 23.10 and later, which otherwise stop it from creating the user namespace it sandboxes commands in. CI applies the same file.

Changed

  • Renamed to OpenSuperBot. Environment variables are OPENSUPERBOT_*, the packages @opensuperbot/*, the bot image opensuperbot/bot:dev, its containers opensuperbot-<id>, the webhook header X-OpenSuperBot-Key and the workflow format opensuperbot.workflow. The start-up templates are ai.opensuperbot.gateway (launchd), opensuperbot.service (systemd) and the OpenSuperBot gateway task (Windows).
  • Nothing breaks for an existing install:
    • OPENGROKBOT_* variables are still read when the new name is unset, and start-up lists the old names it saw.
    • An existing opengrokbot.db stays in use.
    • The old sign-in cookie and the X-OpenGrokBot-Key header are still accepted, and workflows saved as opengrokbot.workflow still load.
    • Docker computers named opengrokbot-<id> are replaced on first use; sign-ins live in the workspace, so they carry over. An image built as opengrokbot/bot:dev is tagged with the new name instead of rebuilt.
    • Running install-task.ps1 again removes the old Windows task, so two gateways never fight over the port.
  • A teammate no longer edits earlier messages when it takes a new screenshot. The older screenshots are swapped for a placeholder when the request is sent, the same as before, except on Claude: Claude rejects later thinking blocks once anything before them changes, so it keeps up to ten (prompt caching makes the repeats cheap) and folds the oldest five away at a time, asking Anthropic's API to drop the thinking blocks that no longer match.
  • PUT /api/settings refuses to change the provider or API base while OPENSUPERBOT_API_KEY pins the key (that key belongs to one service), and drops a saved key when the API base changes without a new one.
  • The default model is now grok-4.7 (was grok-4, which no longer appears in xAI's model list). If you never set OPENSUPERBOT_MODEL, the gateway starts asking xAI for grok-4.7; it reasons by default and costs more per token than the older Grok models. Set OPENSUPERBOT_MODEL to keep a different model.
  • pnpm install no longer compiles better-sqlite3. Its package carries prebuilt binaries for macOS, Linux and Windows and loads them in preference anyway, so the compile was wasted everywhere and, on Windows, failed without Visual Studio's C++ tools. Installing on Windows now needs no compiler.
  • pnpm dev works in PowerShell and cmd: the gateway's dev script set its default UI address with Unix shell syntax, which Windows rejected; the default now lives in gateway/scripts/dev.ts.
  • Every file is checked out with LF line endings (.gitattributes). On Windows git would otherwise convert them, and a bot image built from that checkout fails at entrypoint.sh with /bin/bash^M: bad interpreter.
  • The READMEs and docs give PowerShell versions of the commands and describe Windows local computers.
  • The READMEs describe local computers wherever they still assumed Docker: the FAQ no longer says Docker is required, the computer and security sections say how a local computer differs from a container (files kept to its folder, commands off by default, no resource limits), and the roadmap marks computers without Docker as shipped.

Fixed

  • Turning on commands for a local computer when the sandbox was missing or failed its self-check left the switch off without saying why: the UI treated the gateway's 409 as "already done". The reason now shows in the panel.
  • A machine with no Chrome, Edge or Chromium could not download a browser for its local teammates: the download failed at once with ERR_PACKAGE_PATH_NOT_EXPORTED, because playwright-core does not export its cli.js. The gateway now finds the installer through the package's own folder.
  • The sign-in window of a local computer closed as soon as the teammate used its browser again, or a routine fired — while the operator might be typing a password. It now closes only when the operator speaks again (a message, a decision) or clicks I'm done signing in.
  • The panel's screen froze on the last frame after the browser was reopened for a sign-in; the stream now ends and the panel reconnects to the new browser.
  • The screen viewer no longer retries forever (relaunching the browser each time) after the gateway says there is no screen, and stops streaming while the tab is hidden.
  • The commands switch and the done button lock while the gateway works and show its refusal inline.
  • The shell tool and the briefing describe a local computer's working directory honestly (the folder is the working directory and $HOME; /workspace is the file tools' name for it); ask_for_login tells the model to stop and wait. The node the gateway runs on is first on a sandboxed command's PATH, so nvm-installed nodes work.
  • running() no longer counts a local computer whose browser the operator quit by hand.
  • Commands on a Linux local computer could never be turned on when a runtime manager (~/.cargo/bin, ~/.nvm, …) or the gateway's own folder sat in your home directory, which is the usual case: the sandbox self-check counted bubblewrap's own mount points as a view of your home. It now ignores exactly the folders it binds back, and also checks with a probe file it puts in your home and removes. Found by the new Linux CI job.
  • A gateway test that failed now and then on CI. It waits for two replies in a row, and vi.waitFor gave up at its default of one second on a slow runner; gateway tests now wait up to three seconds, as the web tests already do.

Full changelog: v0.4.0...v0.5.0