Overview
01One start field opens a thread in the chosen project. Under it: everything that needs you grouped by module, the threads that are working, the next scheduled automations, account headroom with the 30-day cost, and what settled today.
The virtual company · self-hosted · early access
A company of agents.
Run by you.
Digital Brain is the workspace for running a company on AI agents instead of headcount — engineering, support, market research, ops. You set the intent in chat, specialized agents do the work end to end, and every turn stays reviewable. On your own hardware, under your own provider accounts.
Engineering
Support
Market research
Ops
Engineering first
In active development right now.
We're building in the open and shipping weekly. Expect progress updates here over the coming weeks, including when early access opens.
Runs on your machine
Loopback-only by default. Pair a phone or another computer over a network you trust.
Bring your own agents
Codex and Claude natively; Antigravity, Cursor, Grok and OpenCode over ACP.
Turns survive restarts
Restart the service and Digital Brain rejoins the exact persisted turn instead of replaying it.
Evidence, not vibes
Required checks keep exit codes, bounded logs and the tested commit. A finished answer proves nothing.
Five modules, one shell
Everything that needs a person, grouped by what it is.
The rail never changes. Each module shares the same list column, header and mono meta strip — so the shape of the work is the only thing that moves.
Threads
02A conversation per piece of work, with its own provider session, Stop control and checkout. Start another thread while agents are running, or open Changes to take an isolated worktree. Files, Changes, Terminal and Preview dock beside the thread.
Automations
03Recurring, event-driven or manual care for every project. Describe a check in plain language; findings arrive as cards you can snooze, dismiss, or authorize with Fix & merge — implementation, validation, review, squash merge and cleanup, one finding at a time.
Reports
04A grouped inbox for what users tell you. Each report carries its assessment, the conversation with the reporter, and a public-reply composer with its own saved draft — so a fix and the answer to the person who reported it stay in one place.
Usage
05Account-first quotas: how much headroom is left, when it resets, what the last 30 days cost. Model and request history, privacy controls and confirmed manual resets sit underneath. The rail's dot means an account is near its limit.
Questions
The honest answers first.
Still missing something? The docs go deeper on install, pairing, providers and required checks.
Is Digital Brain hosted?
No. It runs as a desktop application or a persistent browser service on hardware you own, with a local SQLite database and your own provider sessions. The service stays loopback-only until you bind an address yourself — and you should put TLS in front of it on any network that isn't wholly trusted.
Which agents can it drive?
Codex and Claude Code use their native adapters. Antigravity, Cursor, Grok and OpenCode connect over the Agent Client Protocol. Provider instances are durable and named: each one has its own executable, arguments, model aliases and environment, so a personal and a work identity for the same driver can coexist.
Can an agent merge code without me?
Only when you authorize it. Fix & merge and Implement & merge approve one finding's whole pipeline: an isolated worktree from the latest main, your validation command, a diff review, then a squash-merge request. Your host's repository checks and approvals still gate the merge, and a queued merge can be stopped.
Does a finished turn mean the tests passed?
No — and the interface says so. A provider finishing its response doesn't imply your checks passed. Required checks run after coding turns and keep exit codes, bounded logs, the tested commit and a hash of the actual files and index. Files changing during validation produce a stale result, and missing or truncated evidence stays an explicit limitation.
What happens if the service restarts mid-turn?
With the supervised Codex runtime installed, restarting the app detaches the client but leaves the agent and its turn running; startup rejoins the exact persisted turn, restores pending approvals and hydrates missed messages before accepting input. Claude resumes its saved conversation after a graceful restart. Crashes without a prepared handoff keep the manual recovery controls instead of silently replaying your prompt.
Can I use it from my phone?
Yes. The responsive PWA and the native Android app share the same five modules, tokens and typography. Scan a ten-minute pairing code from Settings → Connections and the workspace opens without typing an address; notifications name the work item and project when an agent finishes, fails or needs input. iOS uses the installed PWA.
What does early access include?
A build you run yourself, the pairing and provider setup guides, and a direct line to the people making it. We onboard in small batches so the feedback loop stays short.
Status
Being built right now, in the open.
Digital Brain is under active development: the self-hosted install path and the first round of automations are being finished now. We'll post updates on this page over the next few weeks — what's shipped, what's next, and when early access opens.
Updates in the coming weeks
Credit · Digital Brain's chat-first interface and local-thread behaviour are inspired by T3 Code. We kept what made it good — a thread per piece of work, a shared checkout, one Stop control — and built durable agent execution, required checks and automations around it.