Protocol-Lattice/harness-router
Organization repository on GitHub
A decision layer embedded into the harness tool-selection loop
- Forks
- 6
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Tools & CLIs
Harness Router
MIT-licensed Python tool router for coding agents that keeps obvious calls on a fast path, asks Jev (default typesafe/jev-1.13 via OpenRouter) for genuinely ambiguous choices, and can run bounded MCTS when multi-step consequences matter. Ships a CLI, MCP server and hook integrations.
- Stars
- 21
- Language
- Python
- License
- MIT
- Published by
- Community
- Last verified
- Oct 2026
About this project
Choose the next tool inside an existing agent loop. Harness Router takes a goal, the current observation, and a set of available tools, then returns a tool choice, confidence, or a fallback to the host planner. It uses Jev through OpenRouter for discrete decisions, caches reusable choices, and supports optional Monte Carlo Tree Search (MCTS) over simulated outcomes.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
- Platforms
- CLIMCPOpenRouter
- Capabilities
- Open sourceRequires API key
Highlights
- Default provider model is typesafe/jev-1.13 through OpenRouter; hybrid mode accepts a choice at confidence >= 0.85 and falls back to the planner below 0.60.
- Registers MCP tools route and route_mcts; route returns next tool, confidence, fallback flag and fallback reason.
- Hook integrations exist for Codex, Claude Code, ohmypi, Antigravity and DeepSeek Harness; the shared Stop hook has a 15-second default decision age limit.
- Above 24 tools it can choose a category first and then a tool within that category; decisions and traces are stored under .harness-router/sessions/.
- Status is alpha, Python 3.11+, MIT; scripts/install_hook.py merges hook configuration and backs up changed files as *.harness-router.bak.
Watch out
Young alpha-stage project (21 stars) under MIT. Routing needs OPENROUTER_API_KEY, and MCTS runs over states and transitions you supply rather than executing real tools.
Questions about this entry
Short answers drawn from the linked project or page.
It keeps obvious calls on a cached fast path and consults Jev for genuinely ambiguous tool choices. Below the fallback threshold it returns control to the host planner with a reason instead of guessing, and the PreToolUse hook validates stored decisions without another Jev request.
Codex, Claude Code, ohmypi, Antigravity and DeepSeek Harness have hook adapters, and a stdio MCP server exposes route and route_mcts to any MCP host. har and harness-router are aliases of the same CLI.
A choice at confidence 0.85 or above is accepted, and below 0.60 the router falls back to the planner. Between those it is treated as uncertain, and route returns the next tool, its confidence, a fallback flag and a fallback reason so the host knows whether it got a decision or a deferral.
States and transitions you supply, not real tools. route_mcts takes a root state, simulated states and transitions, a search budget and an optional Jev prior, and returns the first tool, confidence, fallback, the principal variation and search diagnostics. It is a separate search path you invoke explicitly, and it does not execute anything.
Install the CLI with optional MCP support through uv, pointing at the git repository, then export OPENROUTER_API_KEY. Python 3.11+, status alpha, MIT licensed. The package exposes har, harness-router and harness-router-mcp, and the executable and key must be present in the environment of the harness process itself.
Per harness, in the repository: Codex uses SessionStart to discover tools, precomputes at UserPromptSubmit and PostToolUse, and requests a re-plan at PreToolUse on a confident mismatch. Claude Code uses the same callbacks, ohmypi routes at before_agent_start and tool_result and exposes the selected tool with setActiveTools(), Antigravity uses PreInvocation and PostToolUse, and DeepSeek Harness uses the dsh-hooks-codex pre/post bridge. The shared Stop hook has a 15-second default decision age limit.
Above 24 tools it can choose a category first and then a tool inside that category, decided by an input-size estimate rather than a fixed threshold. Decisions and traces land under .harness-router/sessions/ as per-session decision JSON, a decisions JSONL log and a predecision cache, with an LRU-ish shared cache bounded per session by an environment variable.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.
FacebookClinicnotes
YouTube: Jev explained for routing and judging, with a comparison against LLMs.
Hacker News: Vercel on using Jev for defined decisions and GPT-6 Astra when the job also has to write or use tools.
flashbrew4 pointsTypeSafe AI Jev vs. GPT-6 AstraRead the thread on Hacker NewsHacker News: ASTGL replay chose local Laya routing over hosted Jev on one Mac Studio setup.
Jmeg8r15 pointsDoes an open-weight decision model beat a hosted one? Jev vs. LayaRead the thread on Hacker NewsMore Jev resources like this
3.4kGitHub stars
Vercel Labs' Unix-like coding agent CLI written in Zig, shipping a typesafe_permission_reviewer builtin that runs the agent's permission decisions through Jev rather than an LLM call. Model-agnostic and embeddable via libfx and ACP.
Jev Tools & CLIs#community#coding-agents#cli
Community
vercel-labs/fxZig
vercel-labs/fxZig0GitHub stars
A Pi coding-agent extension that uses Jev to classify each request and route it to one of three configured models. Supports TypeSafe, Cloudflare AI Gateway, Vercel AI Gateway and OpenRouter backends, with doctor checks, shadow mode and an off-by-default setup.
Jev Tools & CLIs#community#coding-agents#routing
20.4kGitHub stars
Rust coding-agent harness (MIT, ~20k stars) whose crates/jcode-base/src/jev.rs is a typed TypeSafe Jev decision transport, used for memory pruning (memory_jev.rs), browser navigation (browser_jev.rs), and voice routing (voice_intent.rs) with separate provider selectors per purpose.
Jev Tools & CLIs#community#rust#coding-agent
Community
1jehuang/jcodeRust
1jehuang/jcodeRust13GitHub stars
Permission gate extension for the pi coding agent: deterministic allow/deny rules and protected paths settle clear cases first, then gray-zone tool calls get one structured classify() verdict; Jev is an optional native classifier via TypeSafe direct, OpenRouter, or other gateways, failing closed on error.
Jev Tools & CLIs#community#guardrails#permissions
Community
jesset/pi-verdictTypeScript
jesset/pi-verdictTypeScript7GitHub stars
Claude Code plugin whose UserPromptSubmit hook asks Jev a Choice over /effort levels (low, medium, high, max, unclear) plus a Noul on whether a hands-off request has a fuzzy spec, then offers a one-key switch before Claude starts when the tip clears its confidence threshold.
Jev Tools & CLIs#community#claude-code#plugin
3GitHub stars
Evaluation scripts, MCP tools, and a skill router built to measure where jev-1.13.0 fits in agent harnesses, with raw results under eval/.
Jev Benchmarks & Evaluations#community#evaluation#agents
Community
Aitejiu/jev-harness-labPython
Aitejiu/jev-harness-labPythonFrom the Jev guides
Original write-ups on what this entry is for.
The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.
Jev Tools & CLIs#gateway#vercel#cloudflare
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.
Jev Tools & CLIs#classification#email#security
Read article
From the Jev community
Posts from builders shipping with Jev right now.