You have the conn. Conn orchestrates a fleet of Claude Code sessions across multiple projects: fleet roster, shared canon, drift detection, eval suites. One operator, one bridge.
For engineers running Claude Code across multiple codebases: consultancies, platform teams, polyglot solo devs.
You can't win the next two years by buying more agents. You win by running every ship leaner, and by running more ships than one captain used to be able to.
We don't replace Claude Code. We optimise it. Conn adds nothing to the terminal under your cursor. Everything it does (fleet, rules, evals, drift) lives one floor up.
Conn is the layer above: sessions, graph, registry, playbooks, fleet. The work itself is still Claude's, and the terminal pane stays Claude's too.
No theme picker for things that shouldn't be themed. One way to lay out a project. One way to register a tool. One shape for a playbook.
Every project runs leaner. Sharp context, rules that don't rot, tokens that earn their place.
One captain runs the fleet. Telemetry across every ship. Attention only when it's owed.
Sidebar is your fleet and graph. Main is the terminal under your hand. Right rail shows every session's context window and task. Nothing hidden, nothing modal.
“The best part is no part. The best process is no process.”Elon Musk
A submarine has no wasted space. Every inch earns its place. Conn brings that discipline to your agent stack: context, rules, and tokens all min-maxed for the mission you're actually on. The pane you just saw is a stock Claude Code session; the fleet, the graph, the rules, the drift monitor all live one floor up. The bridge never reaches down and edits the crew's console.
Every crew member, visible. Context windows, current task, on-station indicator. No hidden sessions.
Terminals survive reboots, network drops, and window closes. Reattach in <200ms, and your TUIs (editors, log viewers, cluster tools) come back rendered, not a garbled smear.
No telemetry, no sign-in, no cloud round-trip. The daemon runs on your box, under your user.
Every pattern that works on one project becomes the default on the next. The fleet gets sharper the longer you run it: N projects, M crew each, one bridge.
Every CLI session, across every project, in one list. Status pip, kind, label, current ctx. Click to conn.
CLAUDE.md, slash commands, conventions, hooks. Version v2026.04, propagated across the fleet, audit trail per-project.
Define test cases by name. Run on demand or post-message. Pass / fail / skip, with timings, traces, and diffs over time.
Per-session deviation from canon, plotted over the day. Cross 10% and the XO flags it. Re-anchor in place; no restart.
What you spent vs raw Claude. ~40% on average across the beta cohort. Cap any project, any crew, any window.
A squad = crew + a checked-in playbook. Reproducible flows: “review PR,” “audit hooks,” “ship migration.” All of it lives in the repo.
Every time a new skill drops or a new harness gets hyped, the temptation is to rip everything out and start over. Three weeks later, the sonar is shinier but the engine runs on diesel and the crew is seasick.
Conn lets you evaluate the new thing against what you already run. Adopt the part that earns its place. Keep the hull, keep the doctrine, keep hunting.
Three ways to spend more compute without spending more of your own context. Fire off a discovery question and keep your session clean. Race N approaches and pick the winner. Run a squad in isolation, each with its own worktree and brief.
Big research questions bloat your main session with notes you won't keep.
Explore hands the question to a read-only subagent, lets it wander through the code and the graph, then returns a compressed summary. The discovery cost (every file read, every side-quest) stays in the child.
$ conn explore "why does login hit the cache twice on first request?"
"Which approach is cleanest?" is a question you usually have to answer in your head.
Race spawns up to 8 parallel runners on the same task, each in isolation. Every run is persisted: wall time, token use, diff. The winner is picked automatically; the rest are yours to compare.
$ conn race "rewrite session.ts using the new store" --agents 4
Two agents on the same tree is a merge-conflict factory.
Squad spins up named agents in their own git worktrees with a tighter per-worktree context contract. Each pulls its own brief, writes to its own branch. You keep the trunk clean; they keep their context focused.
$ conn squad spawn hotfix --agents 3
The one capability that most separates Conn from tmux + Claude Code. Every project is indexed, and Claude can query it.
Every Conn project has a live index of its code. Files, symbols, imports, test coverage, ADR pointers, all addressable by Claude as typed nodes. Graph-path queries replace full re-reads; blast-radius-aware edits replace guessing.
$ sonar blast src/auth/session.ts ◉ 14 dependents · 3 tests · 2 ADRs · 312ms ▸ scoped context handed to claude [1]
Conn is Claude Code-native; that's why the vocabulary works. Other CLIs are designed for, not committed to. The tool SDK already lets you register your own today.
Crew, conn, xo, sonar. Learn these four and you can run the product. The full naval dictionary is below, on tap when you need it.
Claude Code stays Claude Code. The data is the proof.
| Capability | Web chat | VS Code agents | Conn |
|---|---|---|---|
| Multi-project fleet viewall sessions, all repos, one rail | none | per-window | native |
| Shared canon (rule packs)CLAUDE.md + slash cmds, fleet-synced | manual | per-repo | v2026.04 propagated |
| Project knowledge graphqueryable, cached, blast-radius aware | no | no | first-class |
| Eval suitesnamed cases, history, per-project | no | no | first-class |
| Drift detectionper-session deviation from canon | no | no | XO flags it |
| Token-budget telemetrytracked across all crew, all projects | opaque | per-session | ~40% mean |
| Owns the terminalre-skins / hijacks the CLI | n/a | yes | never |
I had eleven Claude windows open. Now I have one bridge and a fleet I can actually steer. Drift dropped to zero in week two.S. Ozaki · staff platform engineer · cohort 03
Windows installer + MSI. WSL build for Linux-on-Windows operators. The CLI is the same shape across both.
Run a tighter ship. Run a bigger fleet.
Private alpha. Windows 11. Rolling invites: expect a week or two before your daemon comes up.
▸ costNo fee, no card, no trial clock. Built for my own fleet; you're invited because the currency is data, friends, and making something new.