> ## Documentation Index > Fetch the complete documentation index at: https://docs.avibe.bot/llms.txt > Use this file to discover all available pages before exploring further. # Vibey > Avibe's own agent: built in, always on, and ready from the browser Workbench or any chat app once it has a model from Model Hub. # Vibey Vibey is Avibe's own official agent. It is built into Avibe and runs inside it, so there is nothing to install: add a model source to [Model Hub](/concepts/model-hub), give Vibey a model, and it is ready to work. It is always on, listed first among the agents, and reachable from the [Workbench](/concepts/workbench) and every connected chat app — Slack, Discord, Telegram, WeChat, and Lark / Feishu. Vibey can do the work itself, and it can run the platform for you: schedule work, wait for conditions, build pages, and hand parts of a job to Claude Code, Codex, or OpenCode, which keep running unchanged alongside it. ## What Vibey is good at ### Running the platform Vibey knows Avibe from the inside. Every Vibey turn carries the same Avibe guidance and Skills catalog that every agent gets, and Vibey acts on it through `vibe` commands that its command tool runs: * **Harness:** run another Agent now, schedule a Task, wait for a condition with a Watch, and check run history. → [Harness](/concepts/harness) * **Show Pages:** build a live page — a dashboard, a diagram, a report — that you can open from your phone. → [Show Pages](/concepts/show-pages) * **Vaults and Skills:** use credentials by name and load the Skills you have installed. → [Vaults](/concepts/vaults) · [Skills](/concepts/skills) * **Computer Use:** in the Avibe desktop app on macOS, with Computer Use turned on, Vibey can observe and operate desktop apps through `vibe computer`, the same way every agent does. The setup wizard sums up the idea: *Vibey coordinates Claude Code, Codex, and OpenCode.* Ask Vibey to plan a change, have Claude Code build it, Codex review it, and OpenCode test it, and it reports the result back to you. ### Coding | Tool | What it does | | - | - | | `read` | Reads a text file, up to 2,000 lines or 50 KB per call, continuing with an offset; or an image (JPEG, PNG, GIF, WebP). | | `write` | Creates or overwrites a file. | | `edit` | Replaces exact text in a file: several separate edits in one call, or every match with `replaceAll`. | | `bash` / `powershell` | Runs a command in the session's working directory: `bash` on macOS and Linux, PowerShell on Windows. | If your PATH has no `rg`, Avibe adds a managed ripgrep to the PATH agent commands use. **Project instructions.** At the start of every turn, Vibey loads your instruction files: `~/.avibe/AGENTS.md` for every project, then the `AGENTS.md` (or `CLAUDE.md`) of each directory from the filesystem root down to the working directory. An `AGENTS.override.md` takes the place of a directory's `AGENTS.md`, and where two files conflict, the more specific one wins. An edit applies on the next turn. Together the files are capped at 64 KiB; past that, Vibey is told which file was cut or left out. Files in directories below the working directory are not loaded up front: Vibey is told that a subdirectory may hold a more specific instruction file, and to read it before working with files there. Edit `~/.avibe/AGENTS.md` with **Global prompts** on the Agents page. ### Working in parallel Independent calls in one response — reads, searches, separate commands — run at the same time. Edits and writes run one at a time, in order. A command still running after 120 seconds, or one Vibey starts in the background on purpose, continues as an Avibe Watch, and Vibey gets a follow-up message when it finishes. `vibe watch list` shows it; removing the Watch ends the command. ## Models Vibey gets every model through Model Hub. It has no Direct mode and no CLI login of its own. * **Sources.** Vibey uses Sources held by the gateway: API keys, and subscriptions signed in as a gateway source. A subscription used natively by its own CLI, such as Claude Code's own login, does not supply Vibey. * **Its model list.** When the first such Source exists, Avibe fills Vibey's list once with the models your built-in Claude Code, Codex, and OpenCode Agents use that those Sources serve. If none qualify, Vibey starts on the first model your Sources list that models.dev describes as taking text and calling tools. Only when no listed model qualifies does the list start empty, and you pick one. * **Picking a model.** In the setup wizard, Vibey's card shows **No model yet. Pick one to start.** with models your Sources supply, plus **All models…**. Later, use **Manage models** on Vibey in **Models > Sources & gateway**, or **Add model** at the end of any of its model lists. Reasoning effort follows the tiers configured for the model. * **Protocols.** Vibey calls each vendor in its native API through the gateway — Anthropic Messages, OpenAI Chat Completions and Responses, and Google Gemini — so routing, fallback, and recovery work as they do for every gateway backend. * **The gateway stays on.** While Vibey has models, the model gateway cannot be turned off: **Vibey's models use the model gateway; remove them before turning it off.** If the gateway is off anyway, a Vibey turn stops with **The model gateway is off, so Vibey has no model to run on. Turn it on in Models, then choose a model.** Model Hub is on by default. If the instance's operator has disabled it, Vibey cannot run, and a new install's first default Agent is an available CLI assistant instead. ## Platforms Vibey runs wherever Avibe runs: macOS, Linux, and native Windows. Native Windows support is new in this release and has so far been tested in CI, not yet across real Windows machines; on Windows, WSL remains the recommended setup. * On macOS and Linux, including WSL, commands run in `bash`. * On native Windows, commands run in PowerShell 7 when `pwsh` is on the PATH, otherwise in Windows PowerShell 5.1, which ships with Windows 10 and 11. Vibey is told which one it has; Windows PowerShell 5.1 has no `&&` or `||`. ## Configure in Avibe 1. Add a model source and give Vibey a model. On a new install, do it in the [setup wizard](/get-started/setup-wizard); Vibey alone can complete setup once it has a model. On an upgraded install, where setup is already complete, add the source in **Models > Sources & gateway**, then use **Manage models** on Vibey. 2. Start a Workbench session with Vibey, or route one or more chat channels to it. 3. Optionally create more Agents on the Vibey backend, each with its own prompt and model. On a new install with Model Hub on, new chats start with Vibey. An upgraded install keeps the default Agent it already had; you can make Vibey the default later with **Set as default** on the Agents page, or `vibe agent default vibey` (use `vibey-2` instead if the built-in Agent was renamed, as described below). When Vibey cannot run, its reply says what is missing instead of handing your message to another Agent: in the Workbench, an admin gets **Open Models**, and anyone else is told to ask an Avibe admin; in a chat app, `/setup vibey` explains where Vibey's models come from. The built-in Vibey Agent's model, reasoning effort, prompt, and description stay editable; only its enable switch is locked on. If you already had an Agent named `vibey`, it keeps its name, and the built-in Agent is named `vibey-2`. ## Sessions, forks, and Stop * **Context.** Vibey manages its own context: when a conversation grows long, it condenses earlier parts without interrupting you. There is no `/compact`. If the conversation still cannot fit, Vibey stops and asks you to start a new session with `/new`. * **Fork.** **Fork session** in the Workbench, or `vibe agent run --fork-session --message "..."`, starts a new Session from Vibey's own transcript through the source's last finished turn; a turn still running is left out. Commands, Watches, and Tasks started before the fork stay with the source Session. * **Stop.** **Stop**, or `/stop` in chat, ends the current turn and any command still running in the foreground. Commands already handed to a Watch keep running until they finish or you remove the Watch. A message you send while Vibey is working is read after its current tool calls finish. * **Restart.** If Avibe restarts mid-turn, the turn ends as interrupted and is not resumed on its own; your next message continues with the full context. Running commands keep running as Watches. ## Limits Vibey is a first version: * **Capabilities:** no MCP client (Avibe capabilities arrive through `vibe` commands); no web fetch, web search, or LSP tools; no per-action approval and no sandbox — Vibey runs its tools as your user without asking first. * **Commands:** no input and no terminal, so prompts and full-screen programs do not work. Pausing or removing a command's Watch ends the command, and `/new` pauses it; such a Watch cannot be resumed or edited. On macOS and Linux, a process that detaches into its own session (for example with `setsid`) is not tracked. Commands inherit the Avibe service's environment, including provider variables such as `OPENAI_API_KEY` set there; the gateway's own token is never in it. * **Output:** no live streaming — text and command output appear when each step completes — and no slash commands of its own. Avibe's chat commands, such as `/new` and `/stop`, work as usual. * **Forks:** only from the latest finished turn, not from an earlier point; a fork's first turn starts without the provider's prompt cache. * **Files and images:** `edit` refuses files over 10 MiB or 200,000 lines, and more than 10,000 `replaceAll` matches. Images are not resized or converted: an image file over about 3.4 MB is left out, BMP is not supported, and a model whose image support is unknown gets a text placeholder. Images Vibey has seen and its context scratch files stay under `~/.avibe`, readable only by your user, and are not deleted with the Session. * **Models:** without known limits, Vibey assumes a 128,000-token context and 8,192 output tokens. Gemini models run through the gateway's conversion: reasoning continuity across tool calls is not kept, a stream without a finish reason fails the turn, a safety-blocked prompt can take until the 10-minute silence limit to fail, and model IDs containing `:` are refused. ## Smoke test ```text theme={null} Read the current project root and summarize the repository in three bullets. ``` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.