Harness Router
Claim this toolRoute AI agent tool calls efficiently, reducing latency and cost
Free, no paid tierVisit WebsiteFreeVisit Website
Tracked since2026
The Bottom Line
- Price
- Free, no paid tier. Compare plans
Key facts
- Routes tool calls before execution, reducing model reasoning overhead.
- Offers both skill-based and automatic hook-based routing modes.
- Integrates with Codex via MCP and supports MCTS for complex decisions.
Pros
- Significantly reduces decision latency (up to 12x faster in benchmarks)
- Lowers cost by using a cheaper model for routing decisions
- Flexible: works as a skill or automatic hook
Cons
- Requires setup and configuration with Codex
- MCTS mode may require tuning for optimal performance
What is Harness Router?
Harness Router is an open-source decision layer for AI agent harnesses, designed to route tool calls before they execute. It works in two modes: a skill that lets the model delegate closed-choice tool selection, or a PreToolUse hook that automatically intercepts and routes calls. The tool uses a compact state representation and a fast routing path for ordinary choices, while ambiguous cases can escalate to a Jev-backed Monte Carlo Tree Search (MCTS) for deeper exploration. It integrates natively with Codex via an MCP server, providing a lightweight stdio server that reuses connections and returns tiny structured results. Harness Router aims to reduce latency and cost by avoiding full model reasoning steps for obvious tool choices, with benchmarks showing significant speedups over baseline model reasoning.
Key Features
- PreToolUse hook for automatic routing
- Skill-based delegation for closed-choice selection
- Native MCP server integration with Codex
- Jev-backed MCTS for ambiguous decisions
- Fast routing with compact state representation
- Open source under MIT license
Pricing
Free
Harness Router is completely free to use with no hidden costs.
Reviews
$99Free with your review
Write a reviewReview Harness Router, get a free AI guide
Share your experience and we will send you Improve Your Thinking Patterns Using ChatGPT, free.
Best Harness Router Alternatives
Top alternatives based on features, pricing, and user needs.
ApifyFreemium
Build, run, and scale web scraping and automation workflows
LangChainFreemium
Build LLM-powered applications
CrewAIFreemium
Multi-agent orchestration framework for autonomous AI teams
DevinPaid
AI software engineer by Cognition Labs
BabyAGIFree
An experimental framework for building self-building autonomous agents with a function management system.
HopliteFreemium
Run coding agents in isolated cloud sandboxes
Still deciding?
Most buyers shortlist 2 or 3 tools before committing. Pull a side-by-side comparison or browse the full alternatives shortlist below.
Explore More
Harness Router FAQ
How does Harness Router decide when to use MCTS versus fast routing?
Harness Router uses fast routing for ordinary, closed-choice tool selections, and escalates to MCTS when there is real ambiguity or when the decision involves a multi-step state graph. The MCTS mode explores side-effect-free state graphs locally, without executing any simulated actions.
Can Harness Router be used with agent harnesses other than Codex?
While the documentation focuses on Codex integration via an MCP server, the underlying routing layer is designed to be harness-agnostic. The PreToolUse hook and skill-based approach can be adapted to other harnesses that support similar extension points.
What is the role of Jev in Harness Router?
Jev is a language model used by Harness Router for routing decisions. It is significantly cheaper than larger models, making it cost-effective for the narrow, closed-choice routing step. Jev can also seed the root of the MCTS search, while subsequent simulations are performed locally.
Does Harness Router execute any tool calls during MCTS simulation?
No. The MCTS mode only explores a supplied side-effect-free state graph. It never executes simulated writes, shell commands, browser mutations, or network mutations. It selects only the first real action to execute.
How is Harness Router installed and configured?
Installation is done via uv tool install with the git URL. Configuration involves adding an MCP server entry to the Codex config file (~/.codex/config.toml), specifying the command, environment variables, and timeouts. After restarting Codex, the MCP tools become available.
What are the MCP tools provided by Harness Router?
Harness Router provides two MCP tools: 'route' for ordinary tool routing and 'route_mcts' for MCTS-based routing. Both take a compact goal, latest observation, and a shortlist of candidate tools, and return a tiny structured result indicating the chosen tool.
How does Harness Router ensure routing decisions match what the model would have chosen?
In benchmarks, Harness Router's routing decisions matched Codex's choices 24/24 times for ordinary routing, and achieved 24/24 optimal choices for MCTS. The routing layer validates calls against the discovered tool inventory before execution, ensuring consistency.
What is the cost advantage of using Jev for routing?
Jev's input token price is about 238x cheaper than a leading frontier model, and its output tokens are free. This makes the routing step significantly more cost-effective, especially when the model would otherwise spend a full reasoning step on an obvious tool choice.
Source: harness-router.vercel.app