Describe a goal. Solix plans it, dispatches one agent per task, and holds every risky command those agents run until you approve it — with the exact line in front of you. Give a task a branch and it runs in its own git worktree. Give the plan a cap and it pauses before spending past it.
Claude Code sessions you start yourself show up here too — on the map, with their cost — and Solix leaves them to run.
All of it on your machine. No account, no telemetry.
Then run solix — it wires itself into Claude Code on first run. Or solix demo for a full sandbox in about ten seconds, with nothing to set up.
MIT licensed·Node ≥ 20·Works with any recent Claude Code
Coming next · Codex support
The same view, with OpenAI Codex sessions running alongside Claude Code under the same gate and the same budget.
See what ships today
// Maestro
Maestro turns one sentence into a dependency-ordered plan — and shows it to you before anything runs. Approve it and Solix dispatches one agent per task, checks each result against its acceptance criteria, retries what fails with feedback, and hands you the diff at the end.
It never gets to skip the gate. Every destructive command still stops and waits for you, and the plan pauses before it spends past its cap — with a warning on the way up, not a surprise at the ceiling.
One sentence. Maestro decomposes it into a dependency-ordered plan of small, self-contained tasks.
Nothing runs until you have read it. Change it, cut tasks, or set a budget cap on the whole thing first.
One agent per task, each result verified against its acceptance criteria, failures retried with feedback. You get the diff.
// Views
The galaxy is the overview — every agent, its status, and what needs you, at a glance. Press V for the List: a sortable table with status, cost, context and health, plus one-click actions. Press it again for Missions, which groups the work by the goal it belongs to.
A map when you're scanning. A table when you're working.
// Not a dashboard
The galaxy is the interface — underneath it is a control plane that holds risky tool calls, isolates each task, and enforces the budget you set. All of it locally.
A real Claude Code hook, not a notification. Every sensitive call from a task Solix launched — a New task or a Maestro plan's agents — is held until you answer. Nothing runs behind your back, and nothing is merely logged after the fact. A `claude` session you start yourself keeps Claude Code's own prompt — Solix watches and costs it, but cannot govern a process it did not start. `solix run` adds a way to type into that session from the browser; it does not change who decides.
Give a task a branch and Solix runs it in a dedicated worktree under ~/.solix/worktrees, so parallel agents never step on each other's files or on the branch you're working in. Same branch, same worktree — reused, not duplicated.
Set a cap on a Maestro plan and it warns on the way up, then pauses before the spend crosses it — resume when you're ready. Sessions Solix launched stop taking new work at their cap.
Solix wraps the session instead of tailing it — so you can send input back into a live agent straight from the browser, not just watch it scroll.
Your existing agent definitions are read in as-is. No second config, no duplicated prompts — the crew you already have shows up in the galaxy.
Store an API key once and grant it per advisor or per plan. A worker only ever sees what it was granted. Values are write-only and encrypted at rest.
// The problem
Running one coding agent is fine. Running four is a different job — and nothing you already have is built for it.
Five terminals open, and no way to tell which one is mid-build, which is waiting on a prompt, and which has been blocked on a permission for twenty minutes.
You can't see an agent approaching its context limit until it starts forgetting what it was doing and quietly producing worse work.
An opus run burns budget while you're in another window. No alert, no ceiling, no record of what the spend bought you.
// Legend
Every part of the view maps to something real.
You / mission control
→ Sun
Claude Code session
→ Planet
Subagent (Task tool)
→ Moon
Built-in advisor
→ Inner crew-ring planet
Skill
→ Asteroid in the belt
Tool call
→ Comet streak
Awaiting permission
→ Red pulsing flare
Context usage
→ Planet size
Model
→ Planet color
Cost vs budget
→ Budget ring
Shared goal
→ Constellation lines
// How it works
solixTwo commands, and the second one is the whole setup. On first run Solix wires itself into Claude Code's hooks so it can see sessions the moment they start — it backs up your settings before merging, prints exactly what it changed, and solix uninstall reverses it. No extension, no account, no cloud project.
Full install guidenpm i -g @shmulikdav/solixsolix# wires the hooks, opens 127.0.0.1:4242Keep launching Claude Code from whatever terminal you like — any number of tabs, any model. Every session shows up on its own, and nothing about your workflow has to change.
The counts stay on screen: how many are running, how many need you, how much they've spent. Press ⌘K to jump to any agent by name, or open Mission Control for the whole picture.
The counts are always on screen. Press ⌘K to jump to any agent by name, any Fleet section, or Mission Control for the whole-workspace view — spend, cost per mission, and what is waiting on you.
// The payoff
Solix hooks into Claude Code before the tool runs, so the whole chain is visible — which agent, which tool, what it wants — and execution stays frozen until you answer.
This is a real Claude Code hook, not a toast you can miss. For a task Solix launched, the tool call is held open while Solix waits — so "deny" means the command never ran, not that it ran and you got told about it. If Solix goes down mid-decision, the call is denied rather than let through.
// Cost
Click the sun — or press W — for the whole workspace on one screen: total spend, cost per completed mission, how many times you had to step in, how much context the fleet is burning, and exactly which agents are waiting on you.
Every number is live and local. Nothing is sampled, aggregated upstream, or shown to anyone but you.
// Features
Every sensitive tool call from a task Solix launched is held, with the exact command or diff in front of you before it runs. Approve or deny in one keystroke, having actually read it. Never a blind yes/no. A `claude` session you start yourself keeps Claude Code's own prompt — solix level sets how much is held.
Give a planet a cap and it wears a ring of live spend. A Maestro plan warns on the way up, then pauses before the spend crosses the cap — resume when you're ready. A session you started yourself gets the ring and an alert, since Solix can gate its commands but can't halt a process it didn't spawn.
solix gain reads the audit log back: commands the denylist refused on its own, decisions you made, plans that paused before crossing their cap, and the spend held at each pause. Nothing is collected and nothing is sent — --share only formats it so you can paste it, by choice.
solix ledger is what you allowed and when, with the three patterns worth a second look: approvals made late at night, a command waved through again and again, and something you denied and later approved. Exports Markdown and CSV.
solix level sets how much Solix holds for you — observe, standard or strict. It states which one is active on startup, and warns rather than quietly succeeding when you lower it. The containment denylist sits underneath all three and never moves, for everything Solix launches.
solix import backfills the galaxy from Claude Code's own transcripts, so a fresh install doesn't open on nothing. Idempotent, sessions keep their real dates, and subagent transcripts are skipped so their spend isn't counted twice.
Give a task a branch and it runs in its own git worktree, so parallel agents never collide on files or branches.
Solix wraps the session, so you can send input back into a live agent from the browser — not just watch it.
Each planet shows how much context window is left, so you refactor the prompt before it derails.
Recurring tasks on an interval, on an inbound webhook, or on demand. Each one says in plain words what it will cost — an interval routine starts a real agent session and spends money on every run, until you disable it.
Group agents under a goal; related planets link with constellation lines.
Scrub back through what every agent did, when, and why.
Your existing claude agent definitions load in as-is — no second config to maintain.
Export your advisor + skill setup as a JSON galaxy. One command to share.
Auto-syncs with Claude Code's background sessions. Zero config.
Jump between planets, approve, and mute with single keys. Never leave the keyboard.
Click the sun for a live cost + status overview of every agent — spend, cost per mission, interventions, and what needs you.
Your crew, skills, teams, routines, connections and orgs in one place, instead of scattered across config files.
Jump to any surface, any Fleet section, or any agent by name — including the idle ones you'd otherwise have to find in the 3D scene.
One-click advisor + skill + routine bundles — ship-fast build team, quality & security gate, launch & devrel, architecture & data.
Create, edit, version and restore skills in-app, and dry-run one without spending anything. The bundled ones stay read-only.
Store an API key once and grant it per advisor or per plan. A worker only ever sees what it was granted; values are write-only and encrypted at rest.
Orgs keep separate clients' plans and goals apart. Switch the active org from the rail.
// Meet the crew
Solix ships with a crew rather than an empty canvas: fourteen specialists — a product manager, a builder, a security auditor, a reviewer and more — plus Maestro, the conductor that turns a goal into a plan. They watch your fleet and surface the things you'd miss. Five fly by default. The other ten are one click away.
Enabling an advisor is free — it just joins your crew. Advisors only cost tokens when you pin one (an always-on session) or put one to work on a task — and Solix shows the spend right on each card.
Product Manager
Builder
UX/UI Designer
Code Reviewer
Security Auditor
// Plugins
Cost can wait for a browser tab. A held command cannot. Solix ships two Claude Code plugins — one puts the decision above your prompt, the other gives you the advisors with no Solix server at all.
/plugin marketplace add shmulikdav/Solix/plugin install solix-band/plugin install solix-crewPending decisions sit above your prompt, with the command in view and Approve and Deny beside it — never on their own, so you cannot wave through something you have not read. When nothing is waiting it is one quiet line: what the fleet is doing and what it has cost.
It draws nothing at all when Solix is not running, which is right for most sessions — and makes /solix-band status the way to find out why, if you expected it and got nothing.
Fifteen specialist advisors as Claude Code subagents — a code reviewer, a security auditor, an architect, a FinOps analyst, a debugger, a release engineer and more. Plus /solix-review on a diff and /solix-audit on a repository.
This one needs nothing but Claude Code. No server, no account, no hooks — the smallest possible way to try the thinking behind Solix before you install it.
// Why not just…
Fair question. Here's the honest matrix.
| Capability | tmux / iTerm | Claude /agents | Cloud dashboards | Solix |
|---|---|---|---|---|
| Blocks the action until you approve | no | yes | partial | yes on tasks it launches |
| See every agent at once | no windows, not fleet | no | yes | yes |
| Approve from one place | no | no per-session prompts | partial | yes |
| Isolated worktree per task | no | no | partial | yes opt-in per task |
| Send input back to a live session | yes | yes | no | yes via solix run |
| Context % per agent | no | partial | yes | yes |
| Cost / budget caps | no no signal | partial | yes | yes |
| Code stays on your machine | yes | yes | no sent to vendor | yes |
| Works with any Claude Code session | yes | partial | no | yes |
| Free & MIT | yes | yes | no | yes |
// Quick start
npm i -g @shmulikdav/solixsolix# wires hooks on first run → 127.0.0.1:4242Want to look around first?
solix demo seeds a full sandbox — thirty agents across eight projects, live spend, comet streaks and a pending permission waiting on you. No Claude Code required.
solix demoRequires Node ≥ 20. Claude Code is optional — solix demo runs without it; you only need Claude Code to watch your real agents. macOS and Linux — on Windows, run Solix inside WSL.
// FAQ
// Mission control
Ten agents working while you watch the one that matters.
npm i -g @shmulikdav/solixThen run solix demo — a full synthetic galaxy in ~10 seconds, no setup.