The review board for your AI agents.
The problem
You have customer interview notes, a product question and a recommendation to write. An AI draft helps with the words. You still need to inspect its evidence, resolve a second opinion and explain the decisions behind the final document.
Sprout Crew puts the execution plan, review findings and decision history in one Mac workspace. You choose the providers and approve the scope before execution.
The architecture
React presents the brief, editable plan, review board, and exports. Tauri commands call a local Rust core. SQLite records decisions and task history. No hosted application backend or model gateway is required.
- Approval binding: the approved scope binds the plan and action policy. Changes require a new approval. Planning requests themselves consume allowance.
- Peer review: work passes through the Rust router; a different provider is the preferred second opinion. One API provider can perform a labelled second pass.
- Untrusted inputs: fenced source material and transitive-origin types support separation of data from authority. Prompt fencing does not prove immunity to injection. App-owned action decisions come from Rust policy, not model prose.
- Ledger: API dollars, plan-window percentage, and tokens stay separate. Unknown is never zero. Configured request bounds are reserved before spending. Token-only routes do not impose a dollar cap. The repository's 50% live-check stop overrides the proposed 85% threshold.
- Secrets: application API keys live in macOS Keychain under
sproutcrew, are masked, and are retrieved at send time.
Code-level checks cover app-owned actions. This is not an operating-system sandbox for a third-party CLI. Subscription CLI routes remain disabled in the launch candidate; users bring OpenRouter or direct provider API keys. Arbitrary shell and dependency installs stay disabled until isolation is proven.
Code work lands as a diff
API-backed Builder steps draft in separate git worktrees. Exact base and patch hashes bind each review. Stale patches require a new reconciled draft. Dependency installation and generated tests remain blocked until process, filesystem and network isolation are proven.
A recorded run
A recorded walkthrough will be added after the native app capture is reviewed. The current website does not claim a live demonstration.
What the build demonstrates
Sprout Crew is a Mac app that runs a crew of AI agents — Claude, Codex, and others — under supervision. You approve the plan, and the approval is a hash: change anything material and it invalidates itself. Agents review each other's work across providers, untrusted content is kept separate from action authority, and every run carries an honest, three-meter cost ledger that reserves budget before it spends. It exists because the documented failure mode of agents isn't weakness — it's irreversibility without a checkpoint. Sprout Crew is the checkpoint.
This is an implementation case study, not a benchmark of review quality or a claim that every model finding is correct.
Computer use
Screen capture and browser or desktop control are outside this build. The approved route configuration disables computer and browser use. Source-page imports, static previews and later MCP connections do not provide that capability.