Hand off the goal. Review the work.
A desktop app for Mac and Windows where agents use your files, browser and connected tools to get work done. Free during alpha, with usage limits. Bring the AI plan you already pay for.
Latest releasemacOS on Apple SiliconAlso on Windows and Linux
Click around. This demo runs in your browser on sample data, so nothing is signed in, billed or saved.
The stall occurs after an empty tool result. The captain reports point to retryOnEmpty in stream.ts.
The proposed sample fix disables the empty-result retry. Send a prompt below to try Plan, Ask to Edit, or Execute locally.
More than code.
- Build softwareWork on a feature with files, a terminal and a browser, then review the changes.
- Edit videoConnect a video-editing tool such as Rend and ask an agent to work with your footage. Requires a separate tool connection.
- Research and writeRead sources in the browser, compare findings and draft a document for review.
- School and office workTurn your files and connected documents into lesson drafts, meeting notes or reports.
Sign in with the plan you already have
Every project, one window.
Workspaces hold projects, projects hold chats, and every chat has a terminal and browser next to it.
Workflow overview
- SStudioAAcmePPersonalWorkspaces
One for each client, team or side project, with its own sources, chats and working environment.
- ConsolesSstorefrontDdocsProjects
A repo or folder. Its chats, branches and saved commands live there.
- Agents stalling bugSafari checkoutPricing copyChats
One job per chat. Run as many at once as you like.
- Lead agent3 captains5 subagentsAgents
A lead agent plans the work and hands pieces to captains and subagents.
- TerminalBrowser · 3 tabsCompute sourceComputers
Chats can open terminal and browser tabs. Connect a remote workspace or compute source when needed.
- Edited 5 files+275−126Commit & push#214 openedShip
Review a run’s changes, then commit, push and open a pull request from the chat.
Use the plan you already pay for.
Consoles is free. Sign in with Claude, ChatGPT, Cursor, Muse or Copilot and your agents run on that plan, not per-token API billing. API keys and local models work too.
- Mix providersRun the lead on Claude and its captains on ChatGPT, in the same chat.
- See what's leftEach plan's usage and reset time is one click from the composer.
- Your own endpointsPoint Consoles at Ollama, OpenRouter or any compatible server.
Connections
Manage your AI provider connections.
Anthropic Subscription
OpenAI
Muse
You talk to one agent. It runs the team.
Give the lead agent a goal. It breaks the job into pieces, hands each one to a captain, and reports back when they're done.
- Captains own a goalEach one edits, commits and runs its own subagents. One can wait on another while the rest work in parallel.
- The right model for every seatPut captains and subagents on a different model or provider than the lead.
- Chats talk to each otherAgents can message other chats to coordinate shared work, ask a question or report a result.
2 working
Agents · 2
Delegate by role.
A role is a job with limits. Decide what it does, then what it gets: the model, the permissions, the skills and the connections. Every agent in that role works inside them.
- ModelsPin a model and thinking level to a role, or send it to another account.
- PermissionsAllow or block tools per role. Explore and Plan can't edit files.
- SkillsLoad only the skills the job needs.
- ConnectionsGive a role GitHub without giving it everything else.
Reviews changes to auth, payments and secrets.
- Model
- Opus 5.5 · High
- Permissions
- Read and search only
- Skills
- code-review
- Connections
- GitHub
@security-reviewer check the login changes before I push
Make your own agents.
Write the system prompt once, pick a model and tools, then @-mention it in any chat or put it on a team.
- Set it up onceSystem prompt, model defaults, tools, sources, skills and trigger keywords, all in Settings › Agents.
- Tag it inType @ in the composer and pick it from Agents, right next to the built-ins.
- Put it on a teamAdd it to a blend and every chat that uses the blend can hand it work.
Settings › Agents. Changes are shared with the other homepage examples. Nothing is saved to a real workspace.
Built-in and custom subagents. Own prompt, model, tools, sources, and skills per agent.
Built-in Agents
Read-only research and search
Design a read-only implementation plan
Implement scoped tasks
Land completed parallel work
Custom Agents
Model mapping by provider
What model each built-in resolves to per connection. Switch a session's provider and the agent follows.
| Agent | Claude | ChatGPT |
|---|---|---|
| Explore | ||
| Plan | ||
| General Purpose | ||
| Merge |
Build your own lineup.
Save a lead model and a team of agents as a blend, then pick it from the model menu like any other model.
- Pick the teamChoose which agents the lead model can hand work to.
- Bring your own agentsPut the agents you make on a team, next to Explore, Plan and merge.
- Switch mid-chatPick a different blend from the model menu, even halfway through a chat.
Settings › Agents › Blends. The team is a selection, not an ordered workflow. Changes are shared with the other homepage examples. Nothing is saved to a real workspace.
Built-in and custom subagents. Own prompt, model, tools, sources, and skills per agent.
Blends
Saved orchestrator + team presets. Pick one from the model menu to switch a session to that lineup.
Agents check their own work.
They can start your dev server, open the page in a browser tab and click through it before they hand it back. Connect a remote machine when the job needs more.
No commands run on your machine
/workspace/storefront
- A terminal and your dev serversOpen a terminal beside a chat. Agents can start and inspect the project commands saved in consoles.yml.
- A browser with tabsSeveral tabs per chat, profiles that keep logins apart, and one proxy setting that covers them all.
- Machines when they need themConnect Consoles Compute as a source for remote Linux or GPU workloads and persistent files.
- Or your own hardwareRun a whole workspace on a server you own, and test iPhone apps in the Simulator tab.
You choose how much they do alone.
Plan only reads. Ask to Edit checks with you before each change. Execute just does it. When the work is done, review the diff and commit, push or open a PR without leaving the chat.
- Three modes, one shortcut⇧Tab cycles Plan, Ask to Edit and Execute.
- Commit, push, PRReview a run's recorded changes, draft a commit message, then commit, push or open a PR.
- AutomationsRun tasks on a schedule or when something happens, and get updates in Telegram, WhatsApp or Lark.
+275−126·commit-popover
Review the diff before committing.
Everything else it does.
Pick a group and click around. Every demo runs in your browser on sample data.
Send a task to see the model, sources, effort and agent selections in its result.
What is Consoles?
A desktop home for AI agents that do real work: build software, research, write and work with your files and connected tools. A lead agent can delegate tasks and bring back work for you to review. Available for Mac and Windows.
Do I need API keys?
No. Sign in with the Claude, ChatGPT, Cursor, Muse or GitHub Copilot plan you already have. API keys and your own endpoints (Ollama, OpenRouter, anything compatible) work too.
Is Consoles free?
Free during alpha (usage limits apply). Consoles does not charge you during the alpha. Your AI provider's subscription or API charges still apply; usage limits are separate from your AI plan.
Can I mix models?
Yes. The lead agent, captains, subagents and your custom agents can each run on a different model or provider, and you can switch the model of a chat at any time.
How much can agents do on their own?
You choose with three modes: Plan only reads and proposes, Ask to Edit checks with you before each change, and Execute just gets it done. Shift+Tab switches between them.
What are captains?
Agents that own one goal from a bigger job. Each works in its own chat, on its own branch if you like, runs its own subagents and reports back to the lead agent when it's done.
Does it run on Windows or Linux?
The app runs on Macs with Apple Silicon and on Windows 10 and 11. On Windows, shell commands need Git for Windows, and the app walks you through installing it. You can also install the Consoles server on a Linux machine and run a workspace there.
Give your agents a home.
Free during alpha for Mac and Windows. Usage limits apply. Bring your own AI plan.