// V0.3.0 · EARLY ACCESS · MACOS

Your coding agents,
in formation.

Run Claude Code, opencode, and Codex side by side in a tiled terminal cockpit. Then observe a teammate's agent live, hand a task straight into their session, or merge cockpits and code together in real time.

Download for Mac Apple Silicon · macOS 12+ See how it works →
clarence · my-project
Clarence
AGENTIC CODING COCKPIT
Workspaces Friends 2
Workspaces 2 +
my-project
main · ~/git/my-project
api-refactor
feat/rate-limit · ~/git/api

Agents 4
my-project
Vesper claudeworking
Pluto claudeidle
Astrid opencodeworking
Otto claudeidle

Pull Requests 3
#42
Add auth middleware remiminnebo/my-project
#41
Fix rate limiting remiminnebo/api
Vesper claude
Refactoring auth module
├ src/middleware/auth.ts
├ src/middleware/auth.test.ts
└ Running tests… 14/14 passing
Pluto claude
Plan ready
Migrate sessions to Redis,
3 files, est. 12 min.
Proceed?
Astrid opencode
Writing unit tests for UserService
→ src/services/user.test.ts
→ 22 test cases
127 files · ~/git/my-project
Otto claude
$ Review PR #42 and leave feedback
Reading the diff
12 files changed · +340 −87
branch checked out locally
// A NOTE ON PROCESS

The SDLC of old is dead.
Long live the SDLC.

Plan, code, review, ship. The loop survives, it just runs in parallel now. The standup is you, coffee in hand, walking the rows.

Tabs & split panes

Every workspace holds multiple tabs, each its own split-pane layout. ⌘T for a new tab, ⌘1–9 to jump, ⌘D to split. Background tabs pulse when an agent needs input.

Pop-out windows

Detach any agent into its own window, Teams-style. Close it any way you like: the agent folds back into the tab it came from, content already painted.

One-click PR review

Open PRs appear in the sidebar with live CI status. Hit review: Clarence checks out the branch and hands it to a fresh agent with full context.

Voice push-to-talk

Hold a hotkey and talk. Whisper transcribes on-device and types straight into the focused agent. No cloud, no keyboard.

Worktrees per agent

Opt a workspace into isolated git worktrees and every agent gets its own branch and directory. Two agents, two branches, one repo, zero collisions.

Drop anything in

Drag a screenshot onto a pane and its path lands in the agent's prompt. Agents get names like Vesper and Pluto so you always know who's who.

// REAL-TIME COLLABORATION BETA

Collaborate.

Watch a teammate's agent work live, hand a task straight into their session, or merge cockpits and drive together. No screen sharing. No context lost.

OBSERVE

Watch any agent, live.

Request to observe a friend's terminal. Their PTY output streams into a live read-only pane in your cockpit, every token, in real time. Ask once, see everything.

  • Live terminal stream · LIVE badge · read-only
  • Friend accepts or declines with one click
ada's session
LIVE ada · auth-refactor
$ claude
✓ Auth module loaded (14 files)
Rewriting middleware to use JWT...
→ src/middleware/auth.ts
Fixing edge case in token refresh
ada wants you to observe
auth-refactor · 1 pane
Accept Decline
HANDOFF

Pass a task to any agent.

Write a prompt and send it straight into a friend's agent session. Their agent picks it up as the next instruction. No Slack, no copy-paste, no context lost.

  • Prompt delivered straight to their PTY
  • Recipient sees a toast · accept or decline
// HANDOFF · Send task to marcus
The API rate-limiter is breaking under load.
Failing test: src/api/limiter.test.ts
Can you fix the sliding window logic?
⌘↵ to send
CO-PILOT

Merge cockpits. Code together.

Invite a friend to co-pilot. Both of your panes stream bidirectionally into a shared workspace: you see theirs, they see yours, every agent running in parallel.

  • Bidirectional · both sides see both sides
  • Layout syncs live as panes open and close
LIVE co-pilot · ada
LIVE ada · api-server
$ claude
Designing REST endpoints
→ src/api/routes.ts
LIVE ada · tests
$ claude
Writing integration tests
22 passing / 1 failing
LIVE you · frontend
$ claude
Building API client layer
→ src/lib/api.ts
LIVE you · styles
$ opencode -c
Styling dashboard
// CLARENCE GRAPH BETA

Agents that share
understanding.

Most multi-agent tools coordinate at the process level: they know an agent is blocked. Clarence goes deeper. A local knowledge graph lets agents hand off semantic context: decisions made, files owned, approaches tried and failed. Agent B picks up where Agent A left off without rediscovering the same ground.

  • Decisions persist. Agent A records "chose Zustand for state". Agent B reads it before touching the same code.
  • File ownership. The graph tracks which agent owns which file. Clarence warns before two agents collide on the same path.
  • MCP native. Any MCP-capable agent (Claude Code, opencode, Codex) can read and write the graph via a local MCP server.
  • Local first. Postgres + pgvector, runs in Docker on your machine. Nothing leaves.
// CLARENCE GRAPH my-project · 4 agents · 12 nodes
Decision Use Zustand for auth state Vesper · 14m ago
↓ SUPERSEDES
Attempt Redux approach, failed Vesper · 22m ago · ✗
↓ SCOPED TO
File src/store/auth.ts OWNED BY Vesper
Pluto is about to touch src/store/auth.ts, owned by Vesper

Ready to run your agents?

macOS 12+ · Apple Silicon

Download for Mac Apple Silicon · macOS 12+