fable-mode

Fable Mode is an open-source MCP control plane used to constrain the behavior of AI coding agents. Its core goal is not to make agents "smarter," but to require them to pass through a mechanized gatekeeping process before writing to the user's workspace. The official description is: let the agent think deeply first, then produce evidence, and undergo adversarial review, only after which does it have the right to write. These gates are enforced by the engine, not suggestions at the prompt level—the principle is "no timing, no evidence, no write permission."

Main capabilities

The workflow is divided into four stages:

• Think (thinking / time-locked deliberation): the agent cannot write until the timer ends. The session is locked from the very start and cannot be unlocked by "confidence" alone.
• Prove: claims require real evidence, including tool receipts, hashes, and invariants. The agent needs to record evidence and invariants; unlock_execution requests issued early will be rejected until both the authoritative timer and the evidence preconditions are satisfied.
• Attack: a red-team swarm (five-vector red-team swarm) attempts to break the code from multiple directions.
• Write: only after passing the above stages is the workspace unlocked.

The gates that run through all stages include: evidence receipts for claims, a five-vector red-team swarm for code, and a sealed verification record. Users can call get_status at any time to view the current stage, remaining time, evidence count, and lock status. Persistent engineering memory is also provided.

How to use

Fable is released as a single PyPI package, containing the runtime, the stdio MCP server, and the complete Agent Skill in the same package. The installation steps are as follows:

1. Install the MCP server (the official project provides multiple options).
2. If the complete workflow is needed, add the optional Agent Skill.
3. Ask the agent to handle a specific coding task with Fable Mode and choose a time budget.
4. A new session starts with execution locked; after the agent records evidence and invariants, track the status via get_status.

In the project where the agent will work

run the setup command: this command resolves within an isolated uv environmentPin the package version and copy the bundled skill to .agents/skills/fable-mode. Use --dry-run to preview, or --target <dir> to specify a different skill directory. For a persistent installation, use pip.

Two invocation/transport methods:

• Run fable-engine as a stdio server.
• The same package also exposes a direct JSON transport: pipe a single fable_session parameter object to fable-mode call. This command uses JSON Lines, with one fable_session parameter object per input line and one JSON result per output line. The official recommendation is to keep the process open throughout the entire workflow so that the authoritative timer and the session are in the same trusted runtime; a single-line pipe is suitable for one-off check calls. Each uvx command may resolve to an isolated environment; if the sandbox preserves the Python environment across multiple calls, pip install is more appropriate. Session data is persisted in the Fable data directory outside that environment, and can be overridden with FABLE_DATA_DIR.

When to use it

The typical scenario given by the official documentation is: teams or individuals who want an AI coding agent to think before writing code, leave verifiable evidence, and undergo adversarial review. It can run alongside the agent the user is currently using, is not locked to a framework, and communicates via MCP.

Optional features and limitations

• Optional AI evidence adjudicator: Because the evidence in a session is written by an AI agent, Fable can (optionally) request an external review model to audit the evidence before the workspace is unlocked. It uses only the standard library, a single bounded HTTPS call, no local model, and consumes almost no extra memory. It is off by default and fails closed when executed. The official documentation states clearly that it can only raise the cost of forging evidence and cannot guarantee that deception is impossible; the mechanical gate remains the primary authority.
• Environment requirements: Python 3.10+, zero runtime dependencies, PyPI package name fable-engine.
• The official documentation explicitly states: This is not a guarantee of "defect-free code," but a verifiable workflow rather than a guarantee; it is also not a larger prompt, and the locking is enforced by the engine rather than by wording; it is not framework-locked, but instead runs alongside existing agents via MCP.

Other information

The project uses the MIT license and is built by REX. The officialdocumentation covers practical guides, an Agent Skill reference, V1-to-V2 migration, V2 architecture, the experimental System 3, and the optional AI evidence adjudicator. Issues and pull requests are welcome. The legacy install-skill command is still retained as a compatibility alias that only installs the skill; neither fable-engine nor pip install fable-engine itself writes instructions to the workspace. Regarding the rendering and red-team testing data mentioned in the examples (such as "7 renders rejected before final · 2 bugs caught · 10/10 red-team probes passed"), these are presentational descriptions of the project, and the official documentation does not currently provide more detailed information on the specific test conditions and reproduction methods.