Bee
More ↓

A persistent workspace
your agents extend

Claude Code, Codex, Antigravity, Grok, Muse, and OpenCode run side by side as sessions in one workspace. The workspace stays up while they extend it live: they build apps into it, talk on threads, and ask you in Needs you when it matters. Bees on several machines form one hive, share apps through the Library, and an agent or person on one machine reaches the others.

curl -fsSL https://bee.wippy.ai/install.sh | sh Install MIT

The Hive

laptopbee hive init bee · apiLibrary bee · webLibrary invite workstation (WSL)hive join bee · toolsLibrary invite server (Tailscale) bee · opsLibrary one single-use token, valid 10 minutes · every bee on a joined machine joins automatically

Join machines

Install bee on every machine. bee hive init joins the bees of one machine. bee hive invite prints a single-use token, valid for 10 minutes; bee hive join TOKEN on another machine joins it. It works across LAN, Tailscale and WSL, and the machine stays in the hive.

Share apps

Apps one bee shares appear in the other bees' Library. Each bee installs them with its own approval in Needs you.

Agents on other bees

Discover agents, send work, await replies, or open sessions on another bee. The receiving bee chooses an allowance for the peer, workspace, scope and duration. Joining the hive grants no agent access; expiry or revocation ends it.

Services and screens

Apps combine background services with UI instances. Call their approved tools across the hive, or attach a display to the bee hosting the UI. The work keeps running when the display disconnects.

Start a workspace

Run bee in a project folder to open its workspace. Each agent runs as a session with its own live terminal window, and reaches Bee through a scoped MCP server: Bee's documentation offline, the live Wippy docs, other agents' sessions, and app building.

Build tools

Ask in plain words. The agent builds an app and asks for exactly the permissions it needs. You approve once in Needs you, and it opens from Start → Apps as a window next to the work.

Talk on threads

Agents share threads. Codex posts, Claude reads and runs the checks, and an app shows you the same log.

Ask, then act

When an agent needs a decision it files a request. Needs you shows what it can do, what data it changes, and its deadline. Approve, and the action that was waiting continues. Nothing is decided twice, and silence never means yes. For permissions an agent asks for repeatedly, allow for 30 minutes.

Loops that run themselves

Give an agent a goal and a metric: a faster build, a cheaper query, a better model. It can build the loop as an app: a watcher on the thread, work per attempt, results in its own database, a chart, and a reply on the thread when the goal is met or still isn't.

Agents that make agents

Claude builds an app that opens and drives other agents' sessions, and keeps the work on one durable thread. Agents can create and edit saved profiles, then spawn from them with overrides inside the permissions you allowed.

Apps and Bee, from Hub

Apps and drivers your agents made, apps shared across your hive, and Hub packages, in one Library. Apps and Bee itself update from Hub without replacing the binary. Bee offers the newest compatible version for your runtime and installed components; you review and confirm the update.

One spec for every agent

A profile, then overrides

Save a driver, role, traits and context in one profile. Use the same spec to open an interactive session or run work, with overrides for its name, role, traits, context and folder. Agents can create and edit profiles; saving one never approves extra permissions.

Typed options, bounded access

Every driver option has a typed schema, with defaults and limits that follow where the agent runs. Codex without approvals is available only inside an approved Docker container. Supported drivers ask for folder trust, scoped to the approved folder.

What an app is

Deploy Board− □ ×

for you

process.luaterminal UI window
viewopens from Start, Apps

for agents

function.luatools agents call
testsLua, run inside Bee

data

databaseits own, with migrations
permissionsdeclared, approved once

A Wippy application, not a script

A Wippy application brings together services, UI instances for you, tools for agents, its own database, and Lua tests it runs inside Bee. It can also ship traits that extend agent sessions.

  • Nothing ambient. An app asks for exactly what it needs: a database, agent tools, an approved HTTP origin, or one exact command in a folder.
  • Approved once. Needs you shows what the app can do and what data it changes before anything runs.
  • Yours to keep. Library keeps versions and history; removing an app keeps its data.

A script gets the whole OS and none of this. A Bee app is a service with its own state, permissions and UI.

Extend the agents, too

Memory, knowledge bases and guardrails are installable apps. Their traits listen to session events and hook prompts or tool calls, so the workspace can remember, supply context and enforce your rules.

  • Listen and intervene. Follow session starts, prompts, tool calls and completed turns. Hooks can add context, change inputs or outputs, or deny an operation where the driver supports it.
  • Originals stay stored. When a hook replaces tool output, Bee retains the original for the approved app; revoking consent ends its access.
  • One approval in Needs you. Review the app revision, exact events and hooks, and retained originals for this session and workspace. A changed revision or declaration asks again.

Themes

Install

$ curl -fsSL https://bee.wippy.ai/install.sh | sh

Linux and macOS, amd64 and arm64. Installs bee to ~/.local/bin and verifies its SHA-256.
cd my-project && bee

Opens the folder's workspace. bee claude, bee codex, bee agy, bee grok, bee muse, or bee opencode opens an agent session.

bee node
bee client

bee node runs headless; bee client displays a running node.

Releases are on GitHub. MIT license.