OpenAI Codex
Implementation · Independent review
Local orchestration for coding agents
From task to verified change.
Agent Maestro runs a plan one task at a time with the provider CLIs you choose. Inspect validation, independent review, fix attempts, and Git results in one place.
Runs locally · Uses your installed agent CLIs · No silent provider fallback
Actual application interface · Dark appearance · Sample dataOpen image to inspect details ↗
Inside the product
Inspect the controls before a task runs.
These captures come from the current React application in dark appearance with its built-in sample API. They show real interface components with demonstration data, not a completed provider run.
Workflow settings stay explicit.
A preset supplies guidance and visible review defaults. Validation, Git mode, provider, and limits remain separate choices.
Findings have a history.
Review cycles retain verdicts and finding IDs. A reported fix becomes resolved only after a later review accepts it.
Provider readiness is visible.
The Agents page distinguishes configured models from a completed availability check. Opening the page does not run a diagnostic.
Supported implementation providers
Choose the right agent for each supported role.
Claude Code
Implementation · Independent review
GitHub Copilot
Implementation · Independent review
Gemini CLI
Implementation · Independent review
Antigravity
Implementation only*
*Antigravity cannot enforce the required read-only reviewer boundary. Every provider requires a compatible installed CLI, authentication, account/model access, and applicable organization policy.
The delivery path
A task moves through visible gates.
Agent Maestro runs one task at a time. You configure the checks and optional review before execution, then inspect the result and any bounded fix attempts.
01PLANOrdered tasks and explicit provider selection
↓02IMPLEMENTOne provider-bound session for the current task
↓03VALIDATEConfigured repository commands, when enabled
↓04REVIEW AND FIXOptional independent read-only review and bounded remediation
↓05FINALIZEConfigured Git action, then the next task
A review finding can return to the original implementation session, followed by validation and review again. Providers, gates, limits, and Git mode are configurable; approval does not guarantee correctness.
Capabilities
An operating layer for sustained agent work.
More than a launcher: provider identity, session history, validation evidence, review findings, and delivery state stay connected.
Multi-provider execution
Choose explicit provider, model, and supported reasoning settings without silent fallback.
Review / fix / re-review
Return stable findings to the original implementation session within configured limits.
Validation gates
Run configured, allowlisted commands and send failures back for correction.
Versioned workflows
Use Fast, Standard, Strict TDD, Debug, Autonomous, or a visible custom combination.
Plans and artifacts
Decompose work and pass registered, checksummed artifacts between tasks.
Git delivery
Choose no Git action, commit-only, or non-force commit-and-push after successful gates.
Recovery and resume
Resume from explicit boundaries, subject to each provider’s session prerequisites.
Operational visibility
Inspect task history, changed files, token coverage, and best-effort notifications.
Local-first orchestration
Your repository stays in the development environment you select.
Agent Maestro runs locally, manages orchestration state in SQLite, and directs installed coding-agent CLIs against the selected repository.
Technical credibility
Built around explicit contracts.
A Java 21 / Spring Boot application with a React/Vite interface and local SQLite persistence. Provider-neutral adapters normalize execution while preserving provider-specific permissions and resume requirements.
Machine-checkable task outcomes and review envelopes.
Persisted provider, model, attempt, session, review, Git, and usage provenance.
Visible limits for attempts, review cycles, autofix, and token budgets.
Optional plan-wide review followed by final validation.
Video
See Agent Maestro in action.
FAQ
The practical details.
What coding agents are supported?
OpenAI Codex, Claude Code, GitHub Copilot CLI, Gemini CLI, and Antigravity support implementation. Operation depends on compatible installed CLIs, authentication, account/model access, and policy.
Can implementation and review use different agents?
Yes, when both selected providers support those roles.
Does Agent Maestro replace the coding agents?
No. It orchestrates installed provider CLIs; it does not include provider accounts, subscriptions, authentication, models, or credits.
Does it work with local repositories?
Yes. Repository control and orchestration state are local. Selected provider CLIs may send prompts or code context to external services according to their configuration.
What happens when review finds a problem?
Findings can return to the original implementation session for a bounded fix, revalidation, and re-review. A fix claim is not resolved until verified.
Does every provider support independent review?
No. OpenAI Codex, Claude Code, GitHub Copilot, and Gemini CLI have verified read-only review contracts. Antigravity cannot enforce the required boundary.
Bring control to agent-led delivery
See Agent Maestro on your workflow.
Contact us at ygrikk1@gmail.com.