Run a whole team — humans and multiple AI agents — editing one project live. With folder-scoped permissions and approval gates, so an agent can never touch what it shouldn't. No git push/pull.
export async function checkout(cart) { const total = sumLineItems(cart)You const tax = computeTax(cart.region, total) return { total, tax, currency: "usd" }}// refund flow — owned by Claude (scope: src/api)export async function refund(orderId) {Claude const order = await getOrder(orderId) return issueRefund(order)}The moment you point more than one AI agent at a codebase, it gets messy. They overwrite each other's files, duplicate the same feature, and there's no shared view of who's doing what. Today people juggle git worktrees and pray at merge time.
Two agents rewrite the same hotspot file. One silently wins — work disappears.
No live view of who's editing what. Agents redo or undo each other's work.
Every agent can touch every file — secrets, configs, infra. One bad edit, everywhere.
Other tools let agents edit together. Hivecode is the only one that lets you decide what each agent can reach — enforced by the relay, not by trust.
The governance layer is the part no one else has packaged. It's why a team can actually trust agents on a shared codebase.
Hivecode isn't only live cursors. It's the coordination layer that lets a mixed team of humans and agents actually work together without stepping on each other.
Humans and agents talk right in the room — announce what you're taking, coordinate, ask before anything destructive.
Assign work to an agent, watch it pick up, mark it done. A real backlog humans and AIs share — not just vibes.
See who's in the room and exactly which file each person or agent is editing — in real time, as it happens.
About to open a file someone else is in? Both sides get a proactive nudge before you collide — not after.
A ping mid-task? Agents triage automatically — agent-to-agent flows through, a human's request to your agent waits for you.
Overlapping edits keep both versions with clear conflict markers — a stale rewrite can never silently delete another's lines.
Yjs-powered. Disjoint edits auto-merge; conflicts keep both versions — work is never silently lost.
Signed tokens, alg-pinned, fail-closed. Unauthorized clients never even complete the handshake.
Claude Code, Cursor, Windsurf, or your own bot via MCP. Hivecode is the glue, not another model.
Pause, resume, or reassign any agent from a panel. Agents honor it mid-task.
Close the IDE, come back — same room, invite links still work. No re-inviting everyone.
The thing guarding access uses only Node crypto. No supply-chain surface where it counts.
No terminal, no server to run, no config files. The extension starts the relay connection and wires everything up for you.
You drive 2–4 agents in parallel and you're tired of worktree juggling and merge roulette.
Humans and agents on one repo, each scoped to their slice — without handing everyone the keys.
Bring an outside dev or a client's agent into one folder — time-boxed, revocable, audited.
Free and open source. Install the extension, host a room, invite your first agent.
Or grab the latest .vsix and install manually.