ZacharyZhang-NYandClaude Opus 5.5
245ff66336
- Home: the goal title box is two line boxes tall (2.8em), so a third line no longer peeks out. - Chat messages, reply fallbacks and assignment DMs follow the app display language; tasks and documents keep the working language. - OpenAI and GPT-account model lists match OpenAI's current line-up; default gpt-6.1-sol. - Onboarding connectors step takes the operator's Composio and logo.dev keys, then lists the toolkits the Composio project enables; without a key Connectors leaves the sidebar. - Google Antigravity removed end to end (migration 111), with the remote-revocation path only it used. - Office: 2x pixel people with outlines, rim light, walk, typing and blink; dithered real-clock lighting; detailed furniture and floors. - Per-agent model picker lists the provider's whole model list, live when connected. - Chat wake context breaks same-millisecond ties by insertion order; the mid-run inbox dedupes by message id. - The stored "clock" theme reads back instead of falling to "system". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.7 KiB
AgentCompany
A desktop app that runs a company of AI agents. You set a goal. An AI CEO hires a team, splits the goal into tasks, and the agents do the work against real model providers. You watch, chat, and approve.
Electron, React, and SQLite. Company data lives in a local database; prompts go to the model providers you connect.
Run it
Needs Node 24 and pnpm 11 (corepack enable pnpm).
pnpm install
pnpm dev
Keys
The app ships with no keys. Add your own in Settings > Integrations:
| Key | Used for | Get one |
|---|---|---|
| Composio API key | Connectors page and agent connector tools | platform.composio.dev |
| logo.dev publishable key | Connector logos | logo.dev |
Model provider keys and sign-ins live in Settings > Providers. All keys are encrypted with the OS keychain.
Test
pnpm verify:ci
Typecheck, lint, unit, contract, integration, and Electron e2e.
These run against scripted model replies, so they prove routing and state, not real-model behavior. pnpm test:live runs real tasks on a real provider and checks that the model reads the evidence it is given, ignores instructions planted in what it reads, and does not redo finished work. Set AGENTCOMPANY_LIVE_PROVIDER (a catalog key such as openai), AGENTCOMPANY_LIVE_SECRET and AGENTCOMPANY_LIVE_MODEL; it fails without them and spends real tokens.
Build installers
pnpm package:win
pnpm package:mac
pnpm package:linux
Without an Apple Developer ID, build macOS with the ad-hoc config:
pnpm build && pnpm exec electron-builder --mac --publish never --config electron-builder.unsigned.yml
GitHub Actions runs verify:ci on pushes to main and on pull requests, and attaches installers to each published release.
License
MIT