Avibe turns your own machine into a runtime for the coding agents you already use. It drives the official Claude Code, Codex, and OpenCode command-line tools behind one agent registry and exposes them through a browser workbench or through chat apps, so a session started at your desk can be continued from a phone. Code, keys, and agent processes stay on the machine; the avibe.bot service issues identity and a tunnel and does not proxy your data.
The workbench is a windowed workspace rather than a chat box: Files, Terminal, and Editor run as built-in apps beside the conversation, arranged like a desktop with a dock. Agents can also hand back a Show Page — a live web page such as a flowchart, dashboard, diff, report, or small app — that you comment on by pointing at an element or screenshot, after which the agent reworks the page or answers exactly where you pointed.
A layer called the Agent Harness gives agents durable primitives: run, schedule, watch, and inspect. An agent can start background work, wait for a condition, run on a cron or one-off schedule, and report back to a chosen conversation, and each task and watch record shows its trigger, agent and session binding, delivery target, and run history. A Runs view draws the delegation graph of parent and background sessions. Vaults hold secrets that an agent requests and you approve from the browser; vault responses never include the secret value, and passkey-gated Protected custody is described as pre-launch.
Browser and messaging are equal surfaces: Slack, Discord, Telegram, WeChat, and Lark/Feishu reach the same agents and sessions, with a thread acting as a session. Connections go outward through Slack Socket Mode, the Discord gateway, Telegram long polling, WeChat polling, or a Lark WebSocket, so normal chat control needs no public inbound ports, and vibe remote opens a tunnel to the local workbench that browsers reach only after sign-in. Agents can be switched mid-conversation by prefixing a message, or routed per project, and reusable skills are shared across backends.
Command-line entry points cover starting and stopping the service, configuration and diagnostic checks, remote access, agent management, scheduling, watches, run history, and vault operations. The runtime is Python 3.9+, and at least one supported agent CLI must already be present. Recently added capabilities include multi-window apps, Show Page annotations, the Agent Harness, standard vaults, voice-to-text, and a skills manager; an Avibe-native agent, protected vault custody, and a hosted onboarding mode are listed as planned.
Features
- Official agent CLIs: Claude Code, Codex, and OpenCode run behind one registry, with model and reasoning choices per agent
- Windowed workbench: Files, Terminal, and Editor as built-in apps alongside chat, with multiple windows and a dock
- Agent Harness: run, schedule, watch, and inspect primitives so agents work in the background and report back
- Show Pages: agent-generated live pages with anchored comments that the agent can answer or rework in place
- Runs graph: the delegation history of parent and background sessions, with execution detail per node
- Vaults: approve secret requests from the browser; vault responses omit the secret value
- Messaging surfaces: Slack, Discord, Telegram, WeChat, and Lark/Feishu share the same agents and sessions
- No inbound ports: outbound Socket Mode, gateway, WebSocket, or long-polling connections carry chat control
- Remote access: vibe remote publishes the local workbench through a sign-in-gated tunnel with fail-closed auth and host checks
- Per-project routing: projects map to different agent backends, and agents can be switched mid-conversation
- Mobile and voice: installable as a desktop or mobile app, with voice-to-text and push notifications when a job needs you
