Skip to main content
Glama

Fable Mode

Agents that think before they write.

       


Fable Mode is an open-source control plane for AI coding agents.

It makes an agent deliberate, produce evidence, and survive adversarial review before it earns permission to write to your workspace. The gates are mechanical, not prompt advice: no timer, no proof, no write access.

Demo

KERR // ORRERY

One self-contained HTML file. Raw WebGL, zero libraries, zero external assets, and zero build step.

https://github.com/user-attachments/assets/8287bbfe-e3ee-4dcf-ba0f-f9ff22ae79bd

7 renders rejected before final · 2 bugs caught · 10/10 red-team probes passed

Fable Mode overview

https://github.com/user-attachments/assets/27f4f8a2-b1bb-4398-a08c-bc9fd93d69d7

Quick start

A session starts locked. Confidence does not unlock it.

  1. Install the MCP server using one of the options below.

  2. Add the optional Agent Skill if you want the full workflow.

  3. Ask your agent to use Fable Mode for a concrete coding task and choose a time budget.

A new session starts with execution locked:

{
  "action": "create_session",
  "session_name": "demo-refactor",
  "objective": "Refactor the parser without changing public behavior",
  "time_budget_minutes": 2
}

The agent then records evidence and an invariant. An early unlock_execution request is rejected until the authority timer and proof prerequisites pass. Use get_status at any point to see the active phase, remaining time, evidence counts, and lock state.

The same gates guard every phase: evidence receipts for claims, a five-vector red-team swarm for code, and a sealed record of what was verified.

One package, two agent environments

Fable ships as one PyPI package. The same package contains the runtime, stdio MCP server, and complete Agent Skill. Setup is explicit so installing an MCP server never silently activates workspace instructions.

Run setup from the project the agent will work in:

uvx --from fable-engine==1.3.9 fable-mode setup --yes

This resolves the pinned package in an isolated uv environment and copies the bundled skill to .agents/skills/fable-mode. Use --dry-run to preview or --target <dir> for another skill directory. For a persistent install, use:

python -m pip install fable-engine
fable-mode setup --yes

Then choose only the invocation that matches the agent environment.

Native MCP client

Run fable-engine as the stdio server. For example:

// Claude Code: claude mcp add fable-engine -- uvx --from fable-engine==1.3.9 fable-engine
// Cursor or another JSON-configured client:
{
  "mcpServers": {
    "fable-engine": {
      "command": "uvx",
      "args": ["--from", "fable-engine==1.3.9", "fable-engine"]
    }
  }
}

Install MCP server in VS Code

Shell sandbox with internet, no MCP host

The same package exposes a direct JSON transport. Pipe one fable_session argument object to fable-mode call:

printf '%s\n' '{"action":"create_session","session_name":"demo","objective":"Verify this change","time_budget_minutes":2}' \
  | uvx --from fable-engine==1.3.9 fable-mode call

The command uses JSON Lines: one fable_session argument object per input line and one JSON result per output line. Keep that process open for a full workflow so the authority timer and session stay in the same trusted runtime. A one-line pipe is useful for a single inspection call. Each uvx command can resolve an isolated environment; pip install is better when the sandbox keeps a Python environment between calls. Session data persists outside that environment in Fable's data directory (FABLE_DATA_DIR can override it).

Python 3.10+, zero runtime dependencies. Published on PyPI as fable-engine.

Skill activation remains explicit

setup is the unified path. The older install-skill command remains as a compatible alias for skill-only installation. Neither fable-engine nor pip install fable-engine writes instructions into a workspace on its own.

How it works

  1. Think — Time-locked deliberation. The agent cannot write until the timer ends.

  2. Prove — Claims need real evidence (tool receipts, hashes, invariants).

  3. Attack — A red-team swarm tries to break the code.

  4. Write — Only then is the workspace unlocked.

Optional: AI evidence adjudicator

The evidence in a session is written by an AI agent, so Fable can optionally ask an external reviewer model to audit that evidence before the workspace unlocks. Stdlib-only, one bounded HTTPS call, no local model, no extra RAM to speak of. Off by default; fail-closed when enforcing. It raises the cost of fabricated proof - it cannot guarantee deception is impossible, and the mechanical gates stay the primary authority. Setup and honest limits: AI evidence adjudicator.

What it is not

  • Not a claim of flawless code. It is a checkable workflow, not a guarantee.

  • Not a bigger prompt. The locks are enforced by the engine, not by wording.

  • Not a framework lock-in. It speaks MCP and runs beside your current agent.

Docs

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md.


MIT License · Built by REX

Available Tools

14 tools
browser_backGo back in browser historyA

Loads the previous history entry in the active or named browser session without adding a new history entry. Use it for history traversal rather than browser_navigate. If history is already at its first entry, it is a no-op that returns current page status. A navigation failure returns an error and preserves the history index; omitting session_id reuses or creates the active/default session.

Inputs: session_id. Returns: object

browser_clickActivate a safe browser targetA

Activates an href-bearing link or submits a GET form through normal browser navigation. POST and other form methods are refused, and GET forms containing password fields are blocked to prevent credential leakage. Returns the resulting page status or a bounded error without bypassing existing navigation controls.

Inputs: element_id (required), session_id. Returns: object

browser_closeClose browser sessionA
Destructive

Closes and permanently discards one in-memory browser session, including its current document and history; persistent profile cookies remain saved. Use it when that tab is no longer needed. session_id targets a specific session; if omitted, the active session is closed. Returns closed, or not_found without changing another session.

Inputs: session_id. Returns: object

browser_forwardGo forward in browser historyA

Loads the next history entry in the active or named browser session without adding a new history entry. Use it after browser_back rather than browser_navigate. If no forward entry exists, it is a no-op that returns current page status. A navigation failure returns an error and preserves the history index; omitting session_id reuses or creates the active/default session.

Inputs: session_id. Returns: object

browser_navigateNavigate current browser sessionA

Loads url into an existing browser session, replacing its current document and adding a history entry. Use this instead of browser_open when continuing in a known session; if session_id is omitted, the active session is reused or a default session is created. It waits up to timeout seconds and returns the resulting URL, title, viewport, scroll position, and indexed elements; network, timeout, size, and parse failures return an error object without changing committed page state.

Inputs: url (required), timeout, session_id. Returns: object

browser_openA

Opens a URL in the stealth agent browser using persistent local logins/profile (<=20 MB RAM ceiling). Supports local dev servers across all languages.

Inputs: url (required), timeout, session_id.

browser_pressA

Edits supported text controls, moves focus with Tab, or activates links with Enter.

Inputs: key (required), element_id, session_id.

browser_reloadA

Reloads the current page in the browser session.

Inputs: session_id.

browser_screenshotCapture current viewportA
Read-only

Captures one 1280x800 PNG of the current viewport without navigation or scroll changes. Use this for the visible region; use browser_snapshot_layers for multi-viewport coverage. An omitted session_id targets the active/default session. Returns one layer with its offset, dimensions, and base64 PNG data.

Inputs: session_id. Returns: object

browser_scrollA

Scrolls the page vertically by delta pixels.

Inputs: delta_y (required), session_id.

browser_snapshot_layersCapture layered page snapshotsA
Read-only

Captures up to max_layers PNGs from the top of the current document, one 1280x800 viewport per layer, without changing the session scroll position. Use this for multi-viewport page coverage; use browser_screenshot for only the current viewport. An omitted session_id targets the active/default session. Returns layer count, offsets, dimensions, and base64 PNG data; max_layers defaults to 3 and is capped at 10.

Inputs: max_layers, session_id. Returns: object

browser_typeReplace text field valueA

Replaces the full value of a text-like input or textarea identified by element_id; it does not submit the form. Use browser_press for single-key edits or activation. An omitted session_id targets the active/default session. Returns status typed, the element ID, and the supplied text; missing or non-editable elements return an error object and are not changed.

Inputs: text (required), element_id (required), session_id. Returns: object

browser_waitA

Waits for a specified duration in seconds.

Inputs: seconds (required), session_id.

fable_sessionFable session, evidence, and review workflowA
Destructive

Runs one Fable operation selected by action. Use it for Fable sessions, gates, evidence/proofs, review rubrics, checkpoints, compression, design audits, research scrapers, and explicitly experimental System 3 operations; use browser_* tools for web-page navigation and interaction. Start stateful workflows with create_session and a session_name; later session actions reuse that name. Only fields documented for the selected action are read, and missing or invalid inputs return an Error: message without raising an MCP transport error. Status/list/telemetry/get/view/check operations are read-only. Create, log, record, set, advance, checkpoint, restore, track, evaluate, register, compress/accumulate, evolve, generate, and scraper operations can persist Fable state or artifacts. unlock_execution remains blocked until its time and rationale gates pass; apply_auto_update can replace installed Fable files. The result is action-specific human-readable text in result; inspect that text for success or Error: before continuing.

Inputs: action (required), tag, code, name, +112 more. Returns: result

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.1/5.0

    Scored across 14 tools

    Disambiguation4/5

    Most tools have clearly distinct purposes: navigation, clicking, typing, scrolling, screenshots, history, and session management. The only potential confusion is between browser_open and browser_navigate, but their descriptions clearly separate the 'new persistent session' vs 'reuse existing session' use cases.

    Naming Consistency4/5

    The 13 browser tools follow a consistent verb_noun snake_case convention (browser_open, browser_click, browser_screenshot). The lone fable_session tool breaks the pattern by using a noun instead of an action verb, but this deviation is acceptable as it represents a different functional domain.

    Tool Count5/5

    14 tools is well within the ideal 3-15 range for a server with both browser automation and Fable session management. Each tool covers a distinct action, and the count feels neither bloated nor thin for the stated scope.

    Completeness4/5

    The browser surface covers the core navigation lifecycle: open, navigate, reload, back, forward, click, type, press, scroll, wait, screenshot, snapshot, and close. Minor gaps like explicit element inspection and form submission are workable via existing tools, and fable_session centralizes Fable operations.

    Maintenance

    ActivityStale
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      A runtime gate for coding agents. Blocks the tool calls that wreck a repo (force-push main, rm -rf, secret exfiltration, CI wipe) and lets normal build and commit work through. Machine-checked git-branch core (z3); the rest is high-precision heuristics. Tested on 3,790 real CI commands, 0 false blocks.
      1
      MIT
    • A
      license
      A
      quality
      B
      maintenance
      A machine-enforced workflow protocol for AI coding agents that eliminates false progress. Tasks require verifiable evidence to close, while decisions, assumptions, and open questions persist in a versioned reasoning graph for seamless agent handoffs. Designed to keep deep, long-running projects coherent across sessions, agents, and months.
      39
      32
      AGPL 3.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Verifiable GitHub handoffs between AI agents: receipts, read-back, no self-review. Every write is read back, acknowledgements are bound to the exact task version, and unsafe actions like self-review or a token in a comment are blocked.
      64 npm
      Apache 2.0