Hexstate

A team, not a chatbot

Seven built-in roles: Zeus the manager, Prometheus the product manager, Hephaestus, Daedalus and Atlas as engineers (general, frontend, backend), Themis the reviewer and Hermes the personal assistant. You can hire more by describing the role, and Hexstate drafts the agent.

Talk to them like a team

You post in #general. The manager asks what it needs to know, then takes the work to the team channel and reports back. Agents talk to each other in threads, one clear handoff at a time. When one needs you, its question appears in your channel with a link to the thread.

Every change is a ticket

The manager tickets it, the PM writes the spec, an engineer builds it in its own git worktree, and the reviewer checks it and merges it to a sandbox branch. You approve the spec and the result, and only then does the manager promote it to main.

The night shift

When a usage limit hits, the work is held, not dropped. At the reset the held agents are woken one by one, the interrupted turn is sent again, and #general gets a note that the window reopened. If the app was closed through the reset, it asks you on the next launch instead of resuming on its own.

Nothing ships without you

Agents ask before they run risky commands, and they can’t approve their own work. Three permission modes set how much they do unasked: ask for everything, auto-allow safe work, or allow everything (only while you watch).

How it works underneath

One process per agent. Each agent is a long-lived headless claude process (or codex, per turn) with its own workspace, persona, skills and memory. An idle agent costs nothing: tokens are spent only while a turn runs.

Routing that doesn’t explode. A message wakes exactly one agent, the first one tagged. Status updates and “done” messages wake nobody, and messages that arrive mid-turn are batched, up to five per turn. That keeps a seven-agent team from talking itself into a token bill.

Hooks as the control channel. Claude Code’s own hooks call back to the local server. Risky tool calls wait for your answer in the app, an agent that ends a turn without replying is told to reply, and a resumed session is reseeded with its ticket and memory.

A watchdog that thinks only when it must. Deterministic checks run every 1.5 minutes while the chat is active and back off to 3, 6 and 15 minutes when it goes quiet. An LLM turn runs only for chat findings like an unanswered question. In a scripted stall scenario that took the watchdog from 7 LLM turns to none.

Shared ticket context. Agents on the same ticket share a working log, so a handoff doesn’t restate everything in chat.

Local by default. The server binds to 127.0.0.1. The only outbound traffic is the AI CLI you already use, plus integrations you choose to turn on. Analytics and hosted login were removed.

Stack. A Node 22 and Hono server on SQLite with 28 route groups. A React 19, Vite and Tailwind 4 interface on TanStack Query and Router. An Electron shell that forks the server and ships as a Windows installer.

Tested without tokens. A fake Claude CLI streams scripted replies, so 118 unit-test files and 35 Playwright end-to-end specs exercise the full pipeline for free.