WAYMARK_v3_DEMO.mp4
Human oversight as a property of the system, not a property of the agent.
Waymark is MCP middleware that sits between an AI agent and the filesystem. Every write_file and bash call is evaluated against your policy before it touches anything — blocked, held for approval, or logged and allowed. Nothing the agent does is permanent until you say so.
📖 Full documentation → — installation, CLI and API reference, policy engine, approvals, rollback, agent monitor, and philosophy.
Control — policy lives in waymark.config.json, not in a system prompt. An agent cannot forget a rule, be jailbroken past it, or accumulate enough context to override it. Every tool call is evaluated in Node.js process space before the filesystem is touched.
blockedPaths → BLOCK (hard deny — secrets, system files)
requireApproval → PENDING (human must approve before execution)
allowedPaths → ALLOW (execute immediately, log it)
(default) → BLOCK (unknown paths are rejected)
Observe — the action ledger records every tool call to SQLite with decision, reason, output and timestamps. A dashboard runs at localhost:47000, waymark watch gives a terminal view, and the Agent Monitor shows every AI session on the machine as a live table.
Recover — before-snapshots are captured at write time. A single action can be rolled back, or an entire session undone atomically in reverse order. Approving a write knowing you can undo it is categorically different from approving without a net.
Zero install:
cd your-project
npx @way_marks/cli init --yes # config + CLAUDE.md, register MCP
npx @way_marks/cli start # start server, open the dashboardGlobal install with daemon:
npm install -g @way_marks/cli
waymark global-setup # register the MCP entry once, across all hosts
cd your-project
waymark init # project init (idempotent)
waymark daemon start # one daemon at localhost:47000 for all projectsRestart Claude Desktop, or reload your Claude Code session, to pick up the MCP server.
Installation guide → · Quickstart → · Configuration →
| Platform | Status | Setup |
|---|---|---|
| Claude Desktop | ✅ Recommended | waymark init |
| Claude Code | ✅ Recommended | waymark init |
| GitHub Copilot CLI | ✅ Supported | waymark init |
| GitHub Copilot Chat | ⏳ Future | Waiting on GitHub MCP |
| CodeWhisperer, Codeium, others | ⏳ Future | Waiting on MCP adoption |
Runs on Windows, macOS and Linux. Platform guide →
| Tool | Role | What it does |
|---|---|---|
| devloop | Build | Multi-agent dev pipeline — architect → worker → reviewer |
| waymark | Run | Policy enforcement and observability for AI agents |
| teststop | Break | Adversarial scenario testing — acts as a real, impatient user |
| docuflow | Document | Decision-context wiki for AI agents |
Every release is published on the releases page with its notes and links to the npm packages for that version. The changelog on the documentation site carries the same history in one page.
Installing @way_marks/cli always gives you the current release.
This repository is the public home of waymark: usage documentation, release notes and issues. It does not contain the source code — waymark is a commercial product and the implementation is developed privately.
- Install:
npm install -g @way_marks/cli - Documentation: shaifulshabuj.github.io/waymark
- Issues and questions: open an issue here
Waymark is free to use, with no warranty and at your own risk. Redistribution, resale, modification and rights in the source code are not granted. Versions published before 2026-08-27 remain under the MIT Licence. See LICENSE.
© Shaiful Shabuj. All rights reserved.