Loom Conductor
Brief once. The fleet ships it.
Loom conducts six real Claude Code sessions in parallel. One Conductor plans the work, splits it across the fleet, keeps every session on track, and verifies the result, so you can hand it long-horizon work and walk away.
Free and open source · Apache-2.0 · No telemetry · Uses your existing Claude login
The Conductor is bring-your-own-model
How it works
You write one line.
Loom does the rest.
Brief
Type a goal in the bar at the bottom of the window. One line is enough.
Plan
The Conductor, a model you bring, breaks the goal into tasks and decides what runs in parallel.
Conduct
Tasks dispatch across six real Claude Code sessions. Loom answers prompts and recovers stalls.
Verify
Each piece of work is checked before it counts as done. Jump into any terminal at any time.
The fleet
Six real Claude Code sessions.
Not a simulation.
Every fleet terminal is the actual Claude Code TUI running on your machine, with your existing Claude login. No wrappers, no replays, no API key for the fleet.
Unattended by default
Claude Code asks before it edits files or runs commands. A fleet of six would stall on every popup, so Loom presses the safe affirmative the instant a prompt appears.
Self-recovering
Rate limits, stalls, and dropped sessions are detected and recovered automatically. Leave it running overnight.
Always yours
Click any terminal in the grid to watch it work, or type directly into it. Cycle a single session's permission mode any time with shift+tab.
Brief in plain English. The Conductor plans the mission, splits it across six sessions, watches every terminal, and verifies the result before it counts.
Mission control
One conversation drives planning, dispatch, steering, and review across all six terminals.
Live supervision
Stalls recovered, prompts answered, rate limits rotated, work re-checked before it counts as done.
Smart steering
A drifting session gets nudged back to its task, and recovered sessions get their work re-sent automatically.
Guardrails
Auto-accept presses only the safe affirmative, and the secret-path deny-list holds either way.
Account rotation
Pooled Claude accounts rotate on usage limits so the fleet keeps moving through long missions.
The workspace
Everything you need to
inspect what the fleet did.
A full workspace ships in the same window, so you can read, edit, and review without leaving Loom.
A terminal that keeps up
xterm.js with WebGL rendering, native PTYs, multi-tab with background streaming, splits, search, true color.
A real editor
CodeMirror 6 with inline AI autocomplete and edit diffs you accept or reject hunk by hunk.
Git, visualized
Stage, commit, and browse a full git graph of what the fleet shipped.
See what ships
Dev servers are auto-detected and open in an inline web preview tab.
Make it yours
A full theme engine with presets, custom themes, and background images.
Two AI layers
The fleet is Claude.
The Conductor is your call.
The fleet runs the official claude CLI under the hood and reuses your existing Claude login. The Conductor that plans and co-drives the fleet is bring-your-own-key: any major provider, any OpenAI-compatible endpoint, or a local model. Keys live in the OS keychain, never on disk.
Read the docs# the fleet is the official Claude Code CLI
$ curl -fsSL https://claude.ai/install.sh | bash
$ claude # sign in once, no API key needed
# then pick a Conductor in Settings > Models
openai · anthropic · google · xai · groq · cerebras
deepseek · mistral · openrouter · lm studio · mlx · ollamaUp and running
Install the CLI once. Loom finds it.
The in-app walkthrough auto-detects the claude CLI, you sign in with your existing Claude account, and pick a Conductor in Settings. About a minute, start to first goal.
Security
Built like a terminal.
Locked like a vault.
Keychain only
API keys live in the OS keychain through the keyring crate. Never on disk, never in localStorage.
Guarded boundaries
A secret-path deny-list blocks sensitive reads and writes, and all provider traffic flows through an SSRF-guarded Rust proxy.
Authorized workspaces
Every file, git, and shell operation is checked against a workspace authorization registry before it runs.
Coding agents got good. Babysitting them did not. Loom exists so that six of them can work while you do something else.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Ready in a minute.
All builds on GitHub Releases
FAQ
Questions, answered.
Do I need an API key?
Not for the fleet. The six Claude Code sessions reuse your existing Claude login through the official CLI. The Conductor is bring-your-own-key: paste a key from any of nine providers, point it at any OpenAI-compatible endpoint, or run a local model through LM Studio, MLX, or Ollama.
What exactly does auto-accept do?
Claude Code asks before it edits files or runs commands. A six-session fleet would stall on every popup, so Loom watches each terminal and presses the safe affirmative the moment a prompt appears. It is on by default and can be toggled in Settings. You can also cycle a single terminal's permission mode with shift+tab.
Which platforms are supported?
macOS, Linux, and Windows. Loom is built on Tauri 2 and Rust with a React 19 front end, so the installer is around 13 MB and the terminal renders over WebGL. On Windows, each tab can run Local or any installed WSL distro.
Is Loom open source?
Yes. Apache-2.0, developed in the open at github.com/yimingbeckmann/loom-conductor.
Does Loom phone home?
No. There is no telemetry and no account beyond your own Claude login. Model keys are stored in the OS keychain and provider calls go through a local SSRF-guarded proxy.
What happens when a session hits a rate limit or stalls?
Loom detects it and recovers automatically: it waits out rate limits, restarts stalled sessions, and the Conductor re-checks the work before marking the task done.