Avibe runs on your own machine and drives the official Claude Code, Codex, and OpenCode CLIs, so you can steer them from a browser, a phone, or a chat app instead of one terminal window. Code, keys, and agent processes stay local. The hosted avibe.bot service handles sign-in and a secure tunnel for remote access without proxying the workspace (voice transcription is the one exception).
The browser Workbench opens files, an editor, a terminal, and agent-built pages as windows beside the chat. A Model Hub pools subscriptions and API keys (Claude, ChatGPT, Gemini, Kimi, xAI, relays, and self-hosted endpoints) behind one local gateway, so any agent can use any source's models. When a source hits a quota or rate limit before answering, the next source in the route takes the request, and the hub tracks usage priced at official API rates.
An Agent Harness gives agents four durable primitives (run, schedule, watch, and inspect) so work continues after the conversation ends, and a run graph shows which agent started which background session. The same agents answer in Slack, Discord, Telegram, WeChat, and Lark / Feishu over outbound connections, with no public inbound ports for chat control. Avibe positions itself as a layer under agents rather than an agent itself, and compares itself with OpenClaw and Hermes on setup, security, and token overhead.
Features
- Workbench: files, editor, terminal, and agent-built apps open as windows beside the chat in one browser tab
- Official agent CLIs: Claude Code, Codex, and OpenCode behind one Agent registry, each with its own model and reasoning effort
- Model Hub: one local gateway for subscriptions, API keys, relays, and self-hosted endpoints, with models mixed across agents
- Automatic takeover: a source that hits a quota, rate limit, or network failure hands the request to the next source in the route
- Usage tracking: usage trends by model and source, priced at official API rates, plus live quotas for Claude and ChatGPT plans
- Agent Harness: tasks for one-off and recurring schedules, and watches for PRs, CI, files, logs, and external signals
- Run graph: shows who started each background session, where it reports back, and the history behind every node
- Show Pages: agents answer with live dashboards, diagrams, reports, or small apps that you annotate in place and can share
- Chat surfaces: Slack, Discord, Telegram, WeChat, and Lark / Feishu, with thread replies continuing the same session
- Remote access: a personal subdomain opens the Workbench from any browser through a tunnel, behind a fail-closed access policy
- Vaults: agents use secrets by name, and vault responses never include the secret value
- Shared skills: a Skill written once loads the same way in all three agent backends
