Divyansh Sharma
← BACK TO OPEN SOURCE
ShellMIT
View the code on GitHub

Fable Toolkit: Skills and Multi-Engine Orchestration for Claude Code

I built this because most AI agent failures are process failures: stopping halfway with unneeded questions, calling files done without checking, or trusting subagent summaries without proof. Fable Toolkit turns Claude Code into an accountable lead that finishes tasks and verifies results against primary evidence.

Watch the film narrated, with sound: making AI agents finish the job, and prove it.
Read the transcript

This is Fable Toolkit, an open source kit for A.I. coding agents.

Most A.I. agent failures are process failures, not intelligence failures.

They stop halfway to ask questions, call a file done without running it, and trust summaries without proof.

So I built skills that turn Claude Code into a lead that finishes the job, and checks it against real evidence.

It is free and open source, for anyone running A.I. agents.

The problem

Most AI coding agent failures stem from broken processes rather than a lack of intelligence. Agents frequently halt mid-task to ask questions nobody is around to answer, write a file and mark the job finished without running it, or accept optimistic summaries from subagents without checking the underlying code. When multiple AI tools are available, developers must manually decide which model handles which task, coordinate prompts, and piece together the final result without an objective verification standard.

What it does

  • Adds seven specialized skills and four agents directly to Claude Code.
  • Orchestrates multi-engine AI teams combining Claude Code, OpenAI Codex, and Google Antigravity based on available sign-ins.
  • Assigns tasks automatically per engine, such as planning to Claude, coding to Codex, and bulk editing to Antigravity.
  • Enforces a six-step workflow: specification, cold-start dispatch, evidence review, bounded fix rounds, final gate verification, and reporting.
  • Requires primary command outputs to satisfy acceptance criteria instead of accepting agent self-reports.
  • Provides FABLE-METHOD.md as an operating manual for tool calling heuristics, debugging, and definitions of done.

How it works

Fable Toolkit operates as a set of skills loaded by Claude Code. When you trigger the team workflow, Claude checks which CLI (command-line interface) tools (Codex or Antigravity) are installed and signed in, consulting your saved choices in fable-toolkit.conf. It then breaks your request into independent acceptance criteria (specific pass-or-fail conditions) that can be validated with runnable commands, and assigns each part to the engine best suited for the job.

Claude dispatches cold-start briefs (self-contained instruction packages without shared memory) to each engine via background processes, collects primary output evidence, and compares results against the acceptance criteria. If a step fails, it coordinates up to three fix rounds, escalating to stronger model tiers or Claude itself, before applying a mandatory final gate check that tests the completed work end-to-end from a clean state.

Get started

  1. Install in Windows PowerShell

    irm https://raw.githubusercontent.com/ddivyanshsharmaa/fable-toolkit/main/install.ps1 | iex

    Installs skills and agents, backs up existing files, and configures engine access.

  2. Install in macOS, Linux, WSL, or Git Bash

    curl -fsSL https://raw.githubusercontent.com/ddivyanshsharmaa/fable-toolkit/main/install.sh | bash

    One-line installation script for Unix-like environments.

  3. Install via Claude Code plugin (macOS/Linux)

    /plugin marketplace add ddivyanshsharmaa/fable-toolkit
/plugin install fable-toolkit@fable-toolkit

    Installs the toolkit directly from within Claude Code.

  4. Run a team task

    use the team to add CSV export to the reports page, with tests

    Ask Claude Code in plain language to dispatch and verify work across engines.

Requirements

  • Claude Code, signed in with whatever account you use (subscription, API key, or a cloud provider)
  • Optional: Codex CLI and Antigravity CLI, signed in with your own accounts
  • On native Windows, Git for Windows if you want Codex or Antigravity on the team (to provide Git Bash for closed stdin dispatches)
  • Tested in September 2026 on Windows 11 with Claude Code 2.1.257, Codex CLI 0.155.1, and the Antigravity CLI (scripts written for macOS bash 3.2 and Linux)

Questions

Do I need separate API keys or paid services to use Fable Toolkit?

No. Nothing in the toolkit requires an API key, a server, or a paid service of its own. It relies on the tools already installed and signed in on your local machine.

What happens if an engine is signed out or offline during a task?

The toolkit runs without that engine for that session and informs you, while preserving your saved configuration. The engine returns automatically as soon as it is signed in again.

Why is Git for Windows required on native Windows for multi-engine teams?

Both the Codex and Antigravity CLIs wait indefinitely unless standard input (stdin) is closed. PowerShell 5.1 cannot close stdin with redirection, so dispatches must run through Git Bash.

Why does the verification process reject agent self-reports?

Agents often report success without checking the actual output. The toolkit requires Claude to execute settling commands directly and record primary evidence before marking any criterion as satisfied.