Harness Router
An open-source tool-routing decision layer designed specifically for AI coding agents. It employs a three-tier routing mechanism—fast path, Jev engine, and Monte Carlo Tree Search (MCTS)—to optimize routing between agents and tool calls. It supports integration as either a Codex hook or a skill.
Timeline (last 90 days)
The Harness Router project has been open-sourced on GitHub, with the repository at Protocol-Lattice/harness-router.
Harness Router is an open-source project that attempts to solve the AI Agent tool selection problem by introducing an independent decision layer.
Harness Router adopts a framework-agnostic design that is not tied to any specific Agent framework.
The latest version of Harness Router adds support for Codex hooks.
Harness Router's SessionStart hook discovers all available tools once at the beginning of a session.
Harness Router's PreToolUse hook inspects and makes routing decisions for every tool call that is about to occur.
Harness Router's PreToolUse decision logic is: pass through directly if the selection is the same, trigger replanning if a better alternative exists, and adopt a fail-open policy without blocking execution if the router itself fails.
The fail-open design ensures that even if the routing layer encounters a problem, the entire Agent will not stall but will instead fall back to the original execution path.
The complete execution flow of Harness Router is: Codex → PreToolUse → shortlist → Harness Router → route / MCTS → execute.
Harness Router offers two usage modes: invoked as an explicit skill/MCP tool call, or transparently intervening via a hook.
11 more timeline events
All Facts (19)
The Harness Router project has been open-sourced on GitHub, with the repository at Protocol-Lattice/harness-router.
50%UnverifiedHarness Router is an open-source project that attempts to solve the AI Agent tool selection problem by introducing an independent decision layer.
50%UnverifiedHarness Router adopts a framework-agnostic design that is not tied to any specific Agent framework.
50%UnverifiedHarness Router's SessionStart hook discovers all available tools once at the beginning of a session.
50%UnverifiedHarness Router's PreToolUse hook inspects and makes routing decisions for every tool call that is about to occur.
50%UnverifiedHarness Router's PreToolUse decision logic is: pass through directly if the selection is the same, trigger replanning if a better alternative exists, and adopt a fail-open policy without blocking execution if the router itself fails.
50%UnverifiedThe fail-open design ensures that even if the routing layer encounters a problem, the entire Agent will not stall but will instead fall back to the original execution path.
50%UnverifiedThe complete execution flow of Harness Router is: Codex → PreToolUse → shortlist → Harness Router → route / MCTS → execute.
50%UnverifiedHarness Router offers two usage modes: invoked as an explicit skill/MCP tool call, or transparently intervening via a hook.
50%UnverifiedHarness Router aims to reduce four types of overhead: incorrect tool calls, redundant tool descriptions in context, token usage, and reasoning wasted on obvious routing decisions.
50%UnverifiedHarness Router was built by Kamil Mosciszko.
50%UnverifiedHarness Router employs a three-tier hierarchical decision strategy—fast path, Jev, and MCTS—automatically matching the processing path based on decision complexity.
50%UnverifiedFor tool calls whose outcomes are nearly certain, Harness Router takes the fast path to deliver a decision directly, avoiding the consumption of additional reasoning resources for simple problems.
50%UnverifiedJev is the core driving engine of Harness Router, responsible for handling ambiguous and non-obvious tool selection decisions.
50%UnverifiedHarness Router supports escalating to MCTS (Monte Carlo Tree Search) for complex multi-step decision scenarios.
50%UnverifiedHarness Router offers two integration modes: as a Codex hook to automatically route every tool call, or as a skill for the agent to invoke explicitly.
50%UnverifiedThe side-channel interception mode as a hook requires no modification to the agent's main code and only needs to be connected at the configuration layer, reducing intrusion into existing workflows.
50%UnverifiedHarness Router is categorized under three domains: open-source tools, developer tools, and artificial intelligence.
50%UnverifiedHarness Router is an open-source decision layer designed for AI coding agents, inserted between the agent and tool calls to make tool selection more efficient and reliable before execution.
50%