SoloSwarmFor people who run many AI agents at once
Mission control for your agent swarm
SoloSwarm is one local dashboard for every agent session on your machine. See which ones are running, stuck or waiting on you. Approve pushes, deploys and payments from your phone before they happen. Know what each project costs, and stop it when it hits budget.
Open source (Apache-2.0) · Runs locally on Windows, Linux and macOS · Claude Code supported today
Ten agents is a lot to watch
You lose track
Which session finished, which one is waiting for an answer, which one ran out of usage overnight? You find out by clicking through tabs.
Nobody is at the wheel
Agents in auto mode can push to main, deploy, change DNS or call a payments API, and you read about it in the summary afterwards.
The bill is a surprise
Parallel sessions burn tokens fast. Without a per-project view you can't tell which job is worth it, or which one is looping.
What SoloSwarm does
Session board
Every session in one table: title, project (worktrees roll up to their project), branch, model, tool calls, cost and a live status: running, needs approval, waiting for you, usage limit, idle.
Risk-tiered approvals
Every tool call is classified before it runs. Reads and local edits go straight through. Outward or irreversible actions wait for you in the dashboard or on Telegram, and no answer means no. Catastrophic commands are refused outright.
Cost and budgets
API-equivalent cost by project, client, model, day and session. Set a daily or monthly budget; when it's spent, the agents in that project are stopped until you raise it.
Audit timeline
For each session: prompts, tool calls, approvals (who decided, when, from where) and usage-limit hits. It's useful when an agent did something you didn't expect.
git statusrunsgit push, deploy, DNS, curl -X POST, payments, outbound MCP toolsasks you · 10 min timeout = denyrm -rf ~, force-push to main, reading SSH keysrefusedHow it works
- Install the app. It runs a small daemon on
127.0.0.1and reads your existing Claude Code logs, so the board and costs fill in straight away. - Turn on the hooks. One command adds HTTP hooks to Claude Code (backed up and reversible). From then on risky actions wait for your approval.
- Connect Telegram (optional). Approve or deny from your phone with one tap while your agents work overnight.
SoloSwarm only watches: it doesn't proxy your model traffic or change how your agents work. If the app is closed, your agents carry on as normal.
Get the v0.1 preview
We're opening the preview to a small group first. Leave your email and we'll send the download (Windows, Linux, macOS) when your spot is ready. No newsletter, no spam.
FAQ
Which agents does it work with?
Claude Code today, including parallel sessions in git worktrees and their subagents. Codex CLI and n8n are next, then OpenClaw and Hermes.
Does SoloSwarm send my code or prompts anywhere?
No. It reads Claude Code's local session logs and receives hook calls on 127.0.0.1. Nothing leaves your machine unless you turn on Telegram approvals, and then only the approval card (the action and project name) goes to your own bot.
What happens if SoloSwarm isn't running?
Your agents carry on as normal. The hooks fail open, so a stopped dashboard never blocks your work.
Is it a security sandbox?
No. The risk rules catch the common outward and irreversible actions (push, deploy, DNS, payments, outbound messages), which stops most expensive mistakes. A determined agent could still phrase a command the rules don't recognise, so keep using sandboxes and least-privilege credentials.
How is cost calculated if I'm on a subscription?
From the token usage Claude Code records for every API response, priced at list API rates. On a subscription that's the API-equivalent cost of your usage, which is what you need to compare projects and spot runaway sessions.
Is it free?
The local app is open source under Apache-2.0. A paid cloud tier (sync across machines, phone app, team approvals) may come later; the local app stays free.