BigBrain · what was weighed

Built here, after looking around.

Most of this system is home-built, and that was checked rather than assumed. Each card below is one question put to the tools that overlap with it: the options looked at, what was concluded, and when it was last checked. Notes on fit for this setup, not reviews.

Most of this system was built by hand, but only after looking at what already exists. Each card is one question: the tools that were looked at, what was decided, and when it was last checked. They say whether a tool fits this setup, not whether it is good.

Evaluation

BigBrain or an off-the-shelf agent tool

Could an existing agent tool replace the home-built pipeline that files work from Linear and runs it unattended?

Could a ready-made tool do the job of the system that hands out tasks and runs them with nobody watching?

Options

  • CLIProxyAPIA self-hosted proxy that serves CLI subscription sign-ins as OpenAI-, Gemini- and Claude-compatible APIs. Anthropic's Claude Code terms bar routing subscription sign-in through other tools, and accounts here are already picked by measured headroom before a job starts.Lets other apps borrow your AI subscriptions. Anthropic's rules do not allow that for Claude, and this system already picks an account with room before each job.
  • OpenRouterOne pay-per-token API across many models. It cannot reach subscriptions, and per-token billing is ruled out here.One pay-as-you-go service for many AI models. It cannot use flat-rate subscriptions, and pay-as-you-go is ruled out here.
  • T3 CodeA desktop, web and mobile control surface that drives each vendor's own CLI sign-in, so no terms conflict. It is built for steering live sessions; the value here is the unattended pipeline behind them.An app for steering AI coding helpers by hand. It signs in the approved way, but here most work runs with nobody at the screen.
  • ConductorA Mac app running parallel Claude Code, Codex and Cursor agents in isolated workspaces. Worktree isolation already exists here, and what it leaves to a person (filing, dedupe, roles, the budget gate, hand-offs) is what BigBrain automates.A Mac app that runs several AI coding helpers side by side. This system already does that, and it also does the parts Conductor leaves to a person.
  • Vibe KanbanA local kanban board whose cards launch agents in isolated workspaces. Closest in shape, but the board here is Linear, and its maintainers say it is winding down.A task board where each card starts an AI helper. The closest match, but this system's board is a different tracker, and the project is winding down.

Conclusion

None adopted. One vendor's unmodified CLI, with work filed from Linear and run unattended, is a deliberate choice, and each tool either conflicts with it or repeats what is already built.

None of them. The system uses one AI company's own app on purpose, and each tool either breaks that rule or repeats something already built.

Last checked

Evaluation

Multi-provider coding tools or the in-house launcher

Should a multi-provider coding tool replace or sit beside the in-house launcher that picks a model and an account, now that other vendors' CLIs and local models are on the table?

Should a tool that works with many AI companies replace, or sit beside, the in-house program that starts each AI session, now that other companies' AIs and AI that runs on this computer are options?

Options

  • T3 CodeDrives the real Claude Code and Codex CLIs, so agents and skills still load. It steers live sessions and has no unattended queue.Runs the real AI apps, so everything set up here still works. But it is for a person at the screen, not for work that runs alone.
  • OrcaA desktop orchestrator that runs any terminal agent in its own worktree and shows each account's usage. Closest to the launcher's job, but it lives in a desktop app.A desktop app that runs many AI helpers and shows how much each account has left. The closest fit, but it is built around its own desktop app.
  • piIts own minimal agent with print, JSON and RPC modes. It replaces Claude Code rather than running it, so this fleet's agents, skills and hooks would not carry over, and Claude reaches it only through billing that is ruled out here.Its own AI helper rather than a wrapper. Nothing set up here would carry over, and using Claude in it means a kind of billing that is ruled out here.
  • OpenCodeIts own agent, with ChatGPT sign-in. Claude subscription sign-in was removed after a legal request from Anthropic, and it replaces Claude Code rather than running it.Its own AI helper. It no longer allows a Claude subscription, and it would replace the app this system is built on.
  • CrushCharm's own agent. Apart from Charm's own subscription and local models, it runs on API keys, which are ruled out here.Another AI helper. Apart from its maker's own subscription and AI run on this computer, it needs pay-as-you-go access, which is ruled out here.
  • claude-code-routerA local proxy that keeps Claude Code but swaps the model behind it for other APIs, mostly by API key. Ruled out on billing.Keeps the Claude app but swaps in other companies' AI, mostly on pay-as-you-go access. Ruled out for that reason.

Conclusion

Still open. None routes work by each account's usage windows before dispatch, which is what the unattended side needs. The leading proposal, not yet accepted: extend the in-house launcher to start other vendors' own CLIs and local models, borrow Orca's idea of a per-account usage readout, and keep T3 Code or Orca as an optional window.

Not decided yet. None of them checks which account has room before handing out work, which is what this system needs. The current suggestion, not yet agreed, is to teach the in-house program to start other companies' apps too, and keep T3 Code or Orca around as an optional window.

Last checked

Evaluation

Reaching live sessions from a phone

How should the live Claude Code sessions on this machine be reached from a phone, without starting remote access by hand in each one?

How can the AI sessions running on this computer be checked from a phone, without switching that on one session at a time?

Options

  • Claude Code Remote ControlBuilt into the CLI. One setting registers every interactive session for remote use. The process has to keep running, and the setting has been known to reset itself.Built into the Claude app. One setting makes every session reachable from the phone. The session has to stay open, and the setting sometimes switches itself off.
  • OrcaA desktop app with a mobile companion and SSH remote worktrees. It wraps the CLI in its own window, so the terminal setup and custom panes here would likely not carry over.A desktop app with a phone app alongside. It replaces the terminal window, so the custom setup here would probably be lost.
  • T3 Code and T3 ConnectA hosted relay: the phone signs in and picks a machine, with no port forwarding. The strongest option for many machines. New threads default to full access, which has to be changed before remote use.A service that lets a phone pick which computer to connect to. Best when there are several computers. It starts with every permission switched on, which has to be changed first.

Conclusion

Remote Control, on for every session, keeping the terminal and its custom panes as they are. Orca and T3 Connect are revisited if sessions move to a server or several machines, with T3 Connect the leading option then.

The Claude app's own remote setting, switched on for every session, with the terminal left as it is. The other two get another look if the work moves to more than one computer.

Last checked

Decision

Remote Control on for every session

Should every interactive Claude Code session register for remote use by default, rather than one session at a time?

Should every AI session be reachable from the phone automatically, rather than switched on one at a time?

Options

  • Per sessionStart remote access by hand in each session. Nothing changes until someone remembers to.Switch it on in each session by hand, and remember to every time.
  • On for every sessionOne setting in each configuration, read at startup. Every new session is reachable without a step.One setting that makes every new session reachable on its own.
  • A third-party appOrca or T3 Connect, as weighed in the phone-access card above. Deferred until there is more than one machine.Orca or T3 Connect, from the card above. Put off until there is more than one computer.

Conclusion

On for every session. The terminal and its custom panes stay as they are. The setting is watched for resetting itself, and the choice is revisited on a move to a server or several machines.

Switched on for every session, with nothing else changed. It gets checked now and then in case it switches itself off, and looked at again if the work moves to more than one computer.

Last checked

Each card condenses a longer write-up kept in a private notes library. Tool details were accurate on the date in each card and may have changed since.

Each card is a short version of longer private notes. Tools change, so details were right on the date shown and may not be now.