Take 95% of the work off Claude: our T3 Code setup

In late October 2026 we kept hitting the Claude usage limit, while our Antigravity accounts sat almost idle. So we changed the split in T3 Code: Claude became the planner, the orchestrator and the reviewer, and the other agents became its workhorses. This guide gives you the same setup in about an hour, whether or not you write code.

What is T3 Code?

T3 Code is a free, open-source app that controls the AI coding agents you already pay for: Claude Code, Codex, OpenCode, Antigravity and others. You see every agent's work in one window. It also lets Claude hand a task to another agent with a tool named delegate_task, which is what makes orchestration possible.

It was launched in early 2026 by Theo Browne and Julius from T3 Tools, under the MIT license. delegate_task and orchestrator_capabilities (the list of your accounts and models) are documented in the T3 Code orchestrator docs↗ (opens in a new tab).

How do you split work between Claude and the other agents?

Give Claude the work that needs judgment and give the workhorses the work a brief can describe. Claude writes the plan and the brief, reviews every result and handles anything touching production, money or secrets. Antigravity and OpenCode do implementation, tests, refactors, codebase sweeps, drafts and research.

How do you set it up?

Three things make it work. Add your providers to T3 Code. Add several accounts, because each one has its own quota. Then give Claude a written rule: it plans, orchestrates and reviews, and the other agents are its workhorses. Without the rule, Claude keeps doing all the work itself.

1. 1

   Add your providers to T3 Code

   Open T3 Code, go to Settings > Providers and add Claude Code first: it is the orchestrator. Then add Antigravity and OpenCode, and sign each one in.

   Check: each provider shows as ready.

2. 2

   Add several accounts, not one

   Add Antigravity again as a new instance and sign it in with another Google account. Repeat for each account. Every instance is a separate worker with its own quota, so they run tasks at the same time.

   Check: in a Claude session, ask "call orchestrator_capabilities and list the accounts". Every instance shows up.

3. 3

   Tell Claude it is the planner, not the worker

   Claude does not delegate on its own. Put the rule in ~/.claude/CLAUDE.md, the instruction file↗ (opens in a new tab) Claude Code reads at the start of every session. This is our exact text, minus our names.

   Copy

   # Workhorse policy (spend other accounts' tokens, not Claude's)
   Claude orchestrates, plans and reviews. Antigravity accounts and OpenCode do the bulk work.
   Quality is not traded away; only who pays the tokens changes.
   - Claude does: the plan, the brief, the decisions, the review of every result, the final
     verification and anything touching prod, money, secrets or merges.
   - Workhorses do: implementation from a filled brief, codebase sweeps, test writing, refactors,
     doc drafts, research, data gathering: any task a clear brief fully describes.
   - How: delegate_task with an Antigravity account or opencode-go/* as the target.
     Call orchestrator_capabilities first. Spread work across accounts; each has its own quota.
   - Brief like a peer: goal, files, what is ruled out, how to prove it works, what to return.
     A workhorse result is a draft until Claude has reviewed it and run the check.
   - Fall back to a Claude subagent only when every workhorse route fails. Say so in the reply.

   Check: in a new session, ask "who should write the tests for this change?". Claude should name a workhorse.

4. 4

   Test one account before real work

   Send "reply READY" to one Antigravity account with a fast model such as gemini-3.8-flash-high. Never test with Claude or a thinking model: a test spends the same quota as real work.

   Expected result: READY within a few seconds.

   Check: on a quota or capacity error, skip the other accounts and move to OpenCode, then Claude.

5. 5

   Delegate with a full brief

   The brief is all the workhorse knows: the goal, the files, what is ruled out, how to prove it works and what to return.

   Copy

   Goal: add a "last login" column to the admin users table.
   Files: src/admin/users-table.tsx, src/db/users.ts
   Ruled out: no new dependency, no database migration.
   Prove it: npm test -- users-table, then paste the raw terminal output.
   Return: the diff and the raw test output.

   Check: the brief names a command whose output you can compare.

6. 6

   Review every result

   Claude reruns the checks itself and reads the diff against the brief. It never trusts the output the workhorse pasted.

   Check: Claude's test output matches the reported one. If not, the result goes back.

How do you keep quality when cheaper models do the work?

Treat every workhorse result as a draft. Claude reruns every check itself, reads the diff against the brief and checks every number the agent wrote. Cheaper models sometimes report test results they never ran. The review is where Claude's tokens are best spent.

On 4 November 2026, two Gemini builds reported check output the script cannot produce, and one wrote made-up figures on a public page. Both looked finished. Since then, every brief asks for raw terminal output, and every number in new text is checked against a source.

How do you get more Antigravity quota?

Each account's quota depends on its Google AI plan. Google says↗ (opens in a new tab) free accounts get a weekly limit, while Google AI Pro and Ultra subscribers get higher limits that refresh every five hours. We put a Gemini subscription on each workhorse account, which multiplied the savings again.

Full disclosure: AuraStack is our own store. We bought our Google AI Pro 12-month subscriptions there and use them every day for this setup. You can also buy it directly from Google.

What else cuts Claude's token use?

Two tools make each Claude session cheaper. A code map answers "where is this?" without opening whole files, and a command filter shortens terminal output before Claude reads it. Hooks force Claude to use both. Together they saved about 30 million tokens in eight weeks.

Both numbers are the tools' own estimates of tokens not read, not lines on an invoice. graft is the @nanonets/graft↗ (opens in a new tab) package and RTK is rtk-ai/rtk↗ (opens in a new tab).

What did it change?

On a measured build, at least 95% of the token volume ran off Claude. Over two weeks, 93% of delegated tasks went to Antigravity or OpenCode, and Claude's fresh tokens per day fell by half while its output grew 47%. The same Claude budget now buys plans and reviews, and the work runs elsewhere.

Claude's review caught four problems the workers missed, including a test still using old camera angles and a scoring step that would have zoomed 2K images four times.

Merged pull requests (finished code updates) went from 166 to 403 over the same two weeks. A pull request is not a fixed unit of work, so that number overstates the gain. Our own estimate of the real productivity gain is about 10%, and most of it comes from one change: work no longer stops when the Claude limit is reached.

What is the bottleneck now?

Computing power. Each agent reads files, runs tests and builds on a real machine. On our 8-core server, two agents at once pushed the load to about 20. Adding accounts no longer helps until we add machines to run them.

On 12 November, 180 of 295 delegated Antigravity runs failed. Some were Google's own capacity errors; we have not yet split the rest by cause. We now give each machine its own accounts, and we are preparing 20 more Gemini accounts. The machines come first.

What next?

• Claude skills: the ones I use every day
• Run Claude Code on a VPS and drive it from your phone