Avibe · the local-first Agent OS

Avibe is the local Agent OS for Claude Code, Codex, and OpenCode. Your agents get a visual Workbench, durable Harness, multi-agent runs, Show Pages, Vault access, and the same sessions across Web and messaging.

curl -fsSL https://avibe.bot/install.sh | bash
Workbench · Apps
Local runtime · official agent CLIsRunning on your machine
Local runtime
Code, keys, and agent processes stay on your machine.
3 agent CLIs
Claude Code, Codex, and OpenCode behind one runtime.
6 surfaces
Workbench, Slack, Discord, Telegram, WeChat, and Lark.
MIT licensed
Open source, portable, and built around your existing agents.
THE PRODUCT, NOT A PROMISE

Everything an agent needs to work like a colleague.

Not a wrapper around one model. Avibe gives the agents you already use a shared operating surface, durable work, secure access, and a route back to you.

Vaults
Messaging platforms
Mobile Workbench
Durable sessions
Shared Skills
Human checkpoints
01 · Workbench · Apps

Windows, a dock, real files — on your desktop.

Open Files, Terminal, Editor, and the pages your agent creates in one windowed workspace. Watch the work happen, make a precise edit, and keep the conversation beside it.

  • Files, Terminal, and Editor run as built-in apps
  • Multiple windows stay open and arrange like a desktop
  • Every project and session remains within reach
Workbench · Apps
WorkbenchThree live apps, one Agent OS, and the conversation still beside the work.
02 · Harness · Tasks & watches

Work that outlives the conversation.

The Harness turns an instruction into durable work. Schedule a recurring brief, watch CI or a file, run in the background, and return to the exact definition and history later.

  • Tasks for one-off and recurring schedules
  • Watches for PRs, CI, files, logs, and external signals
  • Inspect the trigger, owner, session, delivery, and every run
Harness · Tasks & watches
HarnessThe task definition, next trigger, owning agent, and delivery path stay inspectable.
03 · Agents · Run graph

One agent hands work to another. You watch the graph.

Give each job to the right specialist without losing the line of responsibility. The run graph shows who started whom, which sessions are in the background, and where results report back.

  • Claude Code, Codex, and OpenCode in one Agent registry
  • Background sessions stay available without cluttering your inbox
  • Open any node to inspect its transcript and run history
Agents · Run graph
Run graphEvery branch keeps its owner, session, status, and callback path.
04 · Show Page · Annotations

Point at the thing instead of describing it.

When words are not enough, the agent creates a live page: a dashboard, diagram, report, or small app. Comment on the exact element or screenshot, and continue the discussion in place.

  • Interactive pages instead of walls of explanation
  • Element and screenshot annotations carry exact context
  • Agent replies and revisions appear where you pointed
Show Page · Annotations
AnnotationA precise comment goes back to the same agent with the exact element attached.
ALWAYS WITHIN REACH

Secret values stay out of Vault responses. Reachable from anywhere.

Vaults let agents request named credentials through an explicit approval boundary. Commands must still avoid printing secret-bearing output. Messaging surfaces and the mobile Workbench bring the same sessions back to you wherever a decision is needed.

Messaging platforms
Vault access request
Protected preview · approval flow
Requested secret
PROD_DB_URL
Requested by
pm · migration lane
Agent boundary
Vault responses expose metadata, not the value.
Approve accessDeny

Vaults

Reference Standard secrets by name and approve Agent access without putting values in Vault responses. Protected custody remains pre-launch.

Messaging platforms

Slack, Discord, Telegram, WeChat, and Lark reach the same Agents and sessions as the browser.

Mobile Workbench

Install it on your home screen, receive a push when a job needs you, and make the decision from your phone.

Durable sessions

Each thread is resumable. Background sessions keep their transcript, context, owner, and callback path.

Shared Skills

Keep conventions and workflows in one library and reuse them across projects and Agents.

Human checkpoints

Buttons, prompts, annotations, and approvals let the agent ask for exactly the decision it needs.

LOCAL-FIRST BY DESIGN

Your machine is the boundary.

Avibe runs on your machine. Using avibe.bot for remote access does not move your code, projects, or Agent execution to the cloud.

Runs locally

Your code, project data, keys, and agent processes stay on the machine you control.

Official agents, unchanged

Avibe drives the official Claude Code, Codex, and OpenCode CLIs. No extra reasoning loop sits between you and them.

Remote without inbound ports

Optional remote access uses a secure tunnel and fail-closed sign-in. Normal chat integrations connect outward.

Open source and portable

MIT licensed. Bring your own subscriptions or keys, switch Agents per task, and avoid another agent silo.

One command to install

Running in about three minutes.

Install locally, choose the official agent you already use, then decide whether to stay in the Workbench or take it with you.

01

Install Avibe

Run the open-source installer, then follow the short browser setup. Your projects and agent processes remain on this machine.

02

Choose your agent

Connect Claude Code, Codex, or OpenCode with the subscription or keys you already have. Add another Agent whenever a task needs different strengths.

03

Work from anywhere

Use the local Workbench immediately. When you want remote access, pair an optional *-app.avibe.bot address or connect a chat platform.

curl -fsSL https://avibe.bot/install.sh | bash
FAQ

Straight answers.

Does it run local models?

Local-first here means your code, data, and execution stay on your machine — not the model weights. Avibe drives the agent you configure (Claude Code, Codex, OpenCode); OpenCode can point at local or OpenAI-compatible endpoints if you want inference local too.

Where do my code and data go?

Your code, project data, and Agent processes stay on your machine. Remote access does not change that. Requests to an Agent go to the AI provider you configure. If you use voice dictation, audio and surrounding text pass through avibe.bot to speech and text-processing providers.

Do I have to pay for Avibe?

Avibe is open source and free to run. You bring your own agent subscription or API keys, so you pay your provider directly — no markup, no second subscription.

Which agents and platforms are supported?

Agents: the official Claude Code, Codex, and OpenCode CLIs. Surfaces: a built-in browser Workbench plus Slack, Discord, Telegram, WeChat, and Lark / Feishu.

Is remote access required?

No. Avibe works locally without an avibe.bot account. Turn on remote access only when you want to reach the same Workbench away from your machine.

How is Avibe different from an agent?

Avibe is the operating layer around an agent, not another model or assistant loop. It gives official agents a Workbench, durable Harness, sessions, Show Pages, Vault access, and routes to you across browser and messaging surfaces.

START LOCAL · GO ANYWHERE

Your best agent deserves somewhere to live.

Install Avibe, open the Workbench, and give it real work.

curl -fsSL https://avibe.bot/install.sh | bash
Avibe — the Agent OS for Claude Code, Codex, and OpenCode