iSparto

中文版

---

iSparto is an open-source cross-provider AI dev team for solopreneurs, built on Claude Code + Codex CLI. Claude plans, orchestrates, and reviews every line GPT implements; a zero-context GPT pass independently re-checks the plan against the product spec. Each provider's work is reviewed by the other. One command spins up the whole team; you direct it through the Team Lead.

中文用户 can start from docs/zh/quick-start.md — a Chinese quick-start covering install, first use, and the daily workflow.

iSparto uses a deliberate bilingual strategy: user-facing entries (both READMEs + the Chinese quick-start + CONTRIBUTING.md) are maintained in parallel; framework instructions and reference documentation are English-only as a single source of truth, so AI instruction-following stays stable and non-Chinese-speaking contributors can review the framework. See CLAUDE.md > Documentation Language Convention for the full rationale.

v0.8.0 (released 2026-04-20) changed two runtime expectations. tmux 3.x is now a hard dependency (Independent Reviewer runs via codex exec in a dedicated tmux pane, not a Claude Code sub-agent), and the Codex CLI now serves two roles (Developer for implementation + QA, and Independent Reviewer for cross-provider blind review). Full release notes in CHANGELOG.md.

The core idea — one command, the whole team

Every existing AI coding tool — Cursor, Windsurf, Copilot, Claude Code on its own — puts you in a loop with a single agent. You and it trade messages for every decision, every file, every commit. The whole development cycle runs through one conversation window.

iSparto's central move is to turn that single agent into an Agent Team. One command (/init-isparto or /start-isparto) spins up the whole agent team — six roles in parallel: Team Lead plans and coordinates, Teammate writes code prompts in parallel, Independent Reviewer audits with fresh context, Developer implements via Codex, Doc Engineer keeps documentation synced, Process Observer guards the workflow. You direct the team through the Team Lead; the rest stays out of your way until a decision is actually needed.

The team deliberately spans two providers. A single-vendor agent fleet shares one training distribution — and with it, one set of blind spots; a review by the same model family that wrote the code inherits the failure modes it is supposed to catch. iSparto routes the quality path across the provider boundary in both directions by default: Claude reviews the code GPT writes, and a zero-context GPT pass re-checks Claude's plan against the product spec at Wave boundaries. Separate failure modes, independent review paths.