🔷 AskAway Model

AskAway Model is a decision model for coding agents.

Independent plugin, not affiliated with or endorsed by TypeSafe AI, Inc. JEV (TypeSafe's System One model) is the first platform AskAway Model supports. JEV is a trademark of TypeSafe AI, Inc.

A local service that sends the choices of Claude Code, Claude Desktop and Codex to JEV, TypeSafe's decision engine, and logs every query.

🌐 Website · Installation · Usage · Changelog · Report a bug · Security

• 📌 Status
• 💡 Why it exists
• 🎁 What it delivers
• 📸 Screenshots
• ⚙️ How it works
• 📋 Prerequisites
• 🚀 Installation
• 🤖 Install with an AI agent (spec)
• 🔄 Updating
• 🧹 Uninstall
• 🧭 Usage
• 🎛️ Configuration
• 🖥️ Claude Desktop (Chat mode)
• 💻 Codex
• 🔀 Optional proxy mode
• 📊 Dashboard
• 📚 Documentation
• 🔒 Security and privacy
• 🩺 Troubleshooting / FAQ

📌 Status

Early access, 1.3.0-beta.2 (GitHub pre-release, 2026-10-19). The Claude Code plugin packaging is new and its layout may change until a stable release. The gateway and the install.py path are unchanged. The latest stable release is v1.2.0. Versions follow SemVer; see the Changelog.

💡 Why it exists

Agents decide many things on their own: which approach to follow, which source to read, which next action to take, how to classify or score a batch of items. AskAway Model hands those choices to JEV, TypeSafe's decision engine, through an MCP server called askaway-model. Claude Code, Claude Desktop and Codex call it, and a local gateway logs every query on a dashboard.

It is for people who run these agents and want each explicit choice checked by JEV, with a visible record of what was asked and answered.

What it does not do:

• It does not generate text, prove facts on its own or grant permissions. JEV decides only between options the agent has already raised.
• It does not change the model's inference endpoint. Registering the MCP server adds one more tool; the Claude Code hooks only remind the agent when to call it.

🎁 What it delivers

• Five decision tools over MCP: request_decision for one choice between approaches, and jev_classify, jev_verify, jev_score and jev_rank for batches. After installation, /mcp in Claude Code lists the askaway-model server (Usage).
• Risk sets the bar: each query states its risk. A Choice or Score answer counts only above the threshold for that risk (0.65, 0.80 or 0.90 by default). Below it, the client takes its local alternative and does not repeat the decision (Confidence thresholds).
• The key stays on the gateway: the MCP adapter never stores the TypeSafe key, which lives only in the gateway's environment file (~/.config/master-jev-hook/gateway.env, mode 600).
• A dashboard for every answer: http://127.0.0.1:8795/dashboard shows answered and no-decision counts, requests per hour and confidence against the thresholds, in English or Portuguese, in light or dark (Dashboard).
• Claude Code hooks that never block a session: an error, a timeout or an abstention does not stop the session. Only a command that turns off the gateway can be blocked (Claude Code hooks).
• One installer for three clients: install.py configures Claude Code, Claude Desktop and Codex. It validates the whole plan before any write, is idempotent and backs up what it replaces (Configuration).
• Optional Claude Code plugin: claude plugin install askaway-model@keepglide replaces the installer step for Claude Code only (plugin path).
• Verification without cost: scripts/verify.sh checks the gateway, the installed files and each client's configuration without a paid call (Installation, step 6️⃣).

📸 Screenshots

⚙️ How it works

The package has three pieces:

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗