Run a fleet,
stay in command.
Control Center is the cockpit for multi-agent software development. Spawn and direct autonomous coding agents, watch their work unfold in real time, and review and merge what they ship — from one quiet, well-instrumented deck.
- ◆ceoCEORunning
Launching claude via subscription relay [claude-relay]
chatrelay · 2m - ⬡reviewer-1ReviewerRunning
Reviewing #1642 · debounce workspace switcher
reviewweb-client · 4m - ▣qa-1QABlocked
Awaiting your approval — ship verdict on architecture review
planplatform-metrics · 11m - ◈devops-1DevOpsFailed
analyze step failed · ConcurrencyConflictException
pipelinecore-api · 6m
Install Control Center.
A native desktop build, shipping today on macOS. Windows and Linux are close behind.
Free and open source · auto-updates · macOS 13 or later.
New to Control Center? Start with the quick-start guide, or browse the full documentation.
Spawning ten agents is easy. Knowing which one needs you isn't.
Each agent runs on its own branch, in its own copy-on-write worktree, producing pull requests, logs, costs and messages in parallel. Git, your terminal and the GitHub UI were all built for one person writing one branch — so ten agents at once turns into ten tabs you have to babysit.
The hard part was never any single action. It's holding the whole fleet in view at once, seeing what's blocked or waiting on you, and acting in one or two moves — without losing the thread.
Command the work, not the chaos.
One ticket carries the work from request to merge.
A ticket is the single unit of work the whole fleet shares — vendor-agnostic, synced with Linear both ways, and coupled to the pipeline that delivers it. An execution lock means exactly one agent owns a ticket at a time, so two never grab the same work.
- Bidirectional Linear sync — status, assignee and comments stay in step.
- Coupled to a pipeline run, so a ticket can drive the work end to end.
- An execution lock — one owner at a time, never a double-claim.
- Vendor-agnostic aggregate — bring your own tracker.
Your day and your meetings, on the same deck.
Record a call and the deck transcribes it on-device, separates the speakers, and writes the summary, decisions, and action items. Connect Google Calendar to see your day and turn any event into a recorded, summarized meeting.
- On-device transcription and speaker diarization, then an AI summary with action items and decisions.
- Local and private: audio and transcripts never leave your machine.
- Google Calendar in month, week, and agenda views, with RSVP to invitations.
- A heads-up before each meeting, and record-and-link straight from a calendar event.
Four systems that keep a fleet legible.
Agents, review and pipelines all sit on the same foundation. Each system below exists so an agent starts a run with more context — and you keep more control — than the run before it.
Memory that compounds across runs.
Every workspace keeps facts, policies and domains in one hybrid store — full-text and vector search over a knowledge graph. An agent opens its run with the context the last run earned, gated to exactly what its role is allowed to read.
- Hybrid FTS + vector retrieval over a single graph.
- Facts, policies and domains as first-class nodes.
- Role-gated reads — no agent sees beyond its scope.
Your codebase, read as a graph.
A tree-sitter symbol-and-edge index per repo lets an agent jump from a symbol to its callers, its callees and its blast radius before it edits a line — the same impact view you get when you sit down to review the result.
- Tree-sitter symbol + edge index per repo.
- Callers, callees and impact radius on demand.
- Powers code search across the whole workspace.
Seventy-one tools, one open protocol.
A JSON-RPC server exposes agents, review, the code graph, memory and ticketing to any MCP client. Mode gating and approval prompts sit in front of every call, so a tool can never outrun the policy you set for it.
- 71 tools over JSON-RPC, spoken by any MCP client.
- Mode gating + approval prompts on every call.
- The same surface the app itself is built on.
Secure by construction.
Each conversation runs inside an OS-native sandbox — Seatbelt on macOS, bubblewrap on Linux — over copy-on-write worktrees. Credentials are minted per launch, capability-gated, brokered in memory and revoked on teardown. Agents never mutate your source, and secrets never touch disk.
- Seatbelt / bubblewrap isolation per conversation.
- Per-launch, capability-gated credentials.
- Secrets brokered in memory, never written to disk.