Regente
Claude, Codex, and Gemini on one repo — Regente gives each cooperating agent a lane, blocks claim collisions before they reach main, calls the right agent when work needs a handoff, and keeps announcements visible to the whole fleet. Local-first, cross-tool, pre-merge.
Works with what you run
Coordinate the agents you already have open.
Regente attaches to your running agents through MCP and session hooks. No new runtime, no orchestrator, no rewriting your workflow. It joins, it doesn't spawn — a neutral coordination layer for the vendors you already run.
- Join, don't spawn. It coordinates the sessions you've already started — never replaces your agents.
- Every vendor, one repo. Claude Code, Codex, and Gemini CLI have repo-local adapters today; Cursor and other agent CLIs are MCP/manual attach until native hooks are real.
- Five-second setup. Run
regente install, restart/open agents, and each tool gets a Regente onboarding file or MCP attach path. - Local-first. The engine, state, and event log live in
.regente/inside your repo. No cloud account, nothing leaves your machine.
MCP/manual Cursor · Aider · OpenCode
How it works
Caught before merge — not after the PR.
With hooks installed, the first agent to edit a file auto-claims it, and a second agent's edit is blocked before it lands. Claude Code and Gemini CLI block at edit time via hooks; Codex is guarded at commit; Cursor coordinates over MCP. Give each agent a role and the files it owns, and everyone (you included) can see who holds what at a glance.
- Pre-merge, cross-vendor. Hook-guarded agents are stopped at edit time; Codex is caught at the commit guard.
- Interleaved edits flagged. Touching a file another online agent edited in the last 30 minutes warns, even when nothing is claimed.
- A lane for every role. Roles own file globs, so territory is explicit and shared.
The part nobody else does
It doesn't just block. It hands off.
A lock would tell the second agent “no” and leave it stuck. Regente queues it, tells it exactly what to do, and the instant the first agent releases the file, hands it to the next one in line — automatically, with a note. No idle agents, no clobbered work, no human merging.
- 01 · claim claude
public/index.htmlediting now - 02 · queue #1 codex ⤳ granted on release
- 03 · release claude
public/index.html✓ released - 04 · owned codex
public/index.html✓ auto-handed off
to the blocked agent “public/index.html is held by claude (frontend). You're #1 in the auto-handoff queue — it's granted to you the moment claude releases. Meanwhile, take an unclaimed file in your lane.” A coordination layer, not a lock.
Why Regente
Worktrees just defer the problem.
The industry's answer to parallel agents is to isolate each one in its own git worktree — which only moves the collision to merge time, with no warning, across no shared view.
“Why did two agents touch the same file again?”
Because nothing was watching across them. Regente sees participating agents at once, auto-claims a file for its first editor, and blocks a hook-guarded second editor before the edit lands.
“Where do the conflicts actually get caught?”
At edit time for hook-guarded agents, at commit for Codex. Worktrees defer conflicts to merge; review tools catch them after the PR. Regente sits in the pre-merge gap.
“Which tool is neutral across all my agents?”
None of them. Agent Teams is Claude-only; GitHub is merge-time. Regente is vendor-neutral by design — repo adapters for Claude, Codex, and Gemini today, with MCP/manual attach for Cursor and other CLIs.
The honest comparison
Where the collision actually gets caught.
Worktrees defer it to merge. Agent Teams is one vendor. Review tools catch it after the PR. Regente is the local layer that’s pre-merge and cross-vendor, catching the second editor at edit time for hook-guarded agents and at commit for Codex.
| Regente | Git worktrees | Claude Agent Teams | Merge-time review | |
|---|---|---|---|---|
| Vendor-neutral repo adapters for Claude · Codex · Gemini; MCP/manual attach for others | Yes | Not applicable tool-agnostic, but blind across agents | No Claude-only | Yes |
| Catches conflicts pre-merge at claim time, before the edit lands | Yes at claim | No surfaces at merge | Partial task-level | No after the PR |
| Auto-hands the file to the next agent the moment the holder releases it | Yes | No | No | No |
| One live view of the whole fleet native and manually joined agents, one screen | Yes | No each worktree isolated | Partial Claude agents only | No |
| File-level claims across vendors not just task assignment | Yes | Not applicable | No task-level | Not applicable |
| Joins the agents you already run never spawns or replaces them | Yes | Not applicable | No spawns teammates | Not applicable |
| Cross-vendor record of who changed what the pre-merge system of record | Yes | No | No | Partial at merge |
“Merge-time review” = GitHub PRs, CodeRabbit, Greptile — they run once a pull request exists, after the work diverged. The closest tool in spirit is the open-source MCP Agent Mail — its file leases are advisory by default, with optional enforcement only at commit time (a pre-commit guard), and no live cross-vendor GUI. Regente blocks at claim time, adds live edit guards where the client exposes hooks, uses a Codex pre-commit guard where it does not, and shows the whole fleet in one local view — pre-merge and cross-vendor at once.
New in 0.5
Setup that verifies itself. Claims that cannot outlive agents.
0.5 makes the two ends of fleet life honest. The first open walks you through a wizard whose steps tick only when the live engine confirms them, and one command wires every vendor end to end. At the long tail, a dead or idle session can no longer sit on your files: claims expire, queued waiters expire, and tasks assigned to gone agents go back to todo.
- Verified install.
regente installwrites hooks, MCP, and an agent skill for all four vendors, verifies each with per-tool checkmarks, then waits for your first agent. - First-open wizard. The desktop app gates each setup step on live verification, not on clicking next. Runs once, skippable, self-verifying checklist after.
- Claim expiry. Claims of a holder dark past
claim_expiry_minutes(default 60) are released, orphaned auto-claims are swept, and stale assignments revert. - In-session delivery. Direct messages land inside the recipient's agent session via the hook heartbeat; the dashboard shows not delivered, delivered, or read.
The command center
One engine, five live views.
The desktop app is the same engine your agents drive over MCP and hooks: Overview for the fleet, Claims for ownership and the auto-handoff queue, Messages for the backchannel, Tasks for the plan, assign, and drive board, and Calls for delegation.
FAQ
Questions, answered.
- Is it free?
- Yes. Regente is free and open source under the MIT License. Access is rolling out through the waitlist, so join above and we'll send your invite.
- Does it work with my agents?
- Claude Code, Codex, and Gemini CLI have repo-local adapters today. Cursor and other agent CLIs are MCP/manual attach until native hooks are real.
- Where does my code go?
- Nowhere. Regente is local-first — its coordination data (claims, lanes, history) stays on your machine and never leaves it. Your agents still talk to their own model providers, just as they do today; Regente adds no server and sends your code nowhere.
- Is it a security boundary?
- No. Regente is a cooperative guardrail for agents that load its hooks and MCP tools. It prevents accidental collisions among participating agents, but it is not an OS sandbox and a local process can bypass it.
Early access
Join the waitlist.
Local-first, cross-vendor coordination for your whole agent fleet. Join the waitlist and we'll send your invite when the next build opens.