Claude Code
Reads your spec and turns it into one clear task.
Install one open-source loop. Claude Code can send builds, reviews, and plans to GPT-6 Astra, then check every result before it ships.
Reads your spec and turns it into one clear task.
Opens the route to the selected model.
Builds, reviews, or plans inside the scope Claude set.
gpt-6-astrayour project onlytests before shipInstall the loop once. Claude Code gets a direct, repeatable path to GPT-6 Astra for real software work. Claude defines the task, Astra handles the delegated work, and Claude checks the result against your repository.
The package includes the bridge, the /astra command, a queue, verification, a green gate, tests, and a live model probe. Every part runs inside the project you choose.
Run one check. The loop calls GPT-6 Astra and waits for a known response.
doctor --live
ASTRA_OKgpt-6-astraASTRA_OKThe kit is free. Your Claude and OpenAI access are separate and stay under your own accounts.
Run the installer from the root of the project you want Claude and Astra to work on.
The bridge and verification scripts run locally with Node.
Claude is the orchestrator and exposes the /astra slash command after install.
Codex must be installed, signed in, and able to use gpt-6-astra on your account.
OpenAI lists the model ID as gpt-6-astra and documents non-interactive Codex CLI scripting with codex exec. Availability can vary by account and rollout.
Open Terminal in the target repository, then run this exact command.
npm exec --yes --package=github:lasseflagstad/loop-kit -- loop-install "$PWD" .claude/commands/astra.mdThe slash command Claude uses to launch a bounded handoff.
scripts/loop/astra.mjsThe bridge that pins Astra and sets the allowed write scope.
scripts/loop/verify.mjsThe local check that confirms the kit is complete and configured.
scripts/loop/run-next.mjsThe queue runner for repeatable, auditable jobs.
Check the installation first. Then make a real, read-only call with Astra pinned.
node scripts/loop/verify.mjs node scripts/loop/astra.mjs doctor --live ✓ Codex CLI installed ✓ OpenAI account authenticated ✓ model pin: gpt-6-astra ✓ live response: ASTRA_OK
Claude receives one reusable command with three explicit delegation modes.
Astra can edit files inside the repository. Claude reviews the handback and runs the project gates.
/astra build Add the approved API validationAstra analyzes the supplied scope without writing to the workspace.
/astra review Check this diff for auth regressionsAstra studies the repository and returns a plan for Claude to judge before work begins.
/astra plan Map the migration in five bounded stepsEvery delegated task keeps its scope, proof, and shipping gates attached.
Objective, constraints, acceptance checks, and allowed scope.
Mode and model are explicit in the command.
Edits, findings, or a plan come back through a recorded process.
The orchestrator compares the result with the original job.
Repository evidence makes the final shipping decision.
Claude can call Astra for implementation, independent review, and planning whenever the job needs another strong model.
Give Astra the approved scope and let it implement inside the repository.
/astra buildSend the error, the affected files, and the acceptance check back through the loop.
/astra buildAsk Astra to inspect a risky diff while the workspace stays read-only.
/astra reviewLet Astra study the repository and return a bounded plan before edits begin.
/astra planThe Claude + Astra loop gives two strong models one way to work together. Trust infrastructure can carry the same research and business context into your offers, marketing, sales, and delivery.
See the operating system that keeps specs, builds, reviews, and green gates connected.
Turn a rough idea into a bounded agent build with tests and approval gates.
Design reusable prompts for recurring business work.