$ 01 · Decision layer

Stop reasoning
about obvious
tool calls.

harness-router keeps the easy path cheap. Obvious action? Call the tool. Real ambiguity? Ask Jev. Multi-step consequences? Search a bounded tree.

uvpipsource
$ uv tool install --force --with 'mcp>=2,<3' 'git+https://github.com/Protocol-Lattice/harness-router.git@main'
2 MCP tools
1 Jev prior max in MCTS
0 router calls on obvious steps
MIT open source
01

The fastest route is often no router.

Earlier skill-driven routing paid extra planner turns around Jev and made the complete loop slower. The current design reverses the default: normal tool calling is the fast path. route exists only for real closed-choice ambiguity.

known file→ read

No routing call. The next action is already obvious.

edit finished→ test

No routing call. Deterministic transitions stay local.

4 plausible tools→ route

Compact state in. One tool choice out.

02

Native MCP. One tiny choice.

Codex gets a native stdio server instead of a heavy helper flow. route takes a compact goal, latest observation, and a small shortlist. The server keeps its provider alive, reuses connections, and returns a tiny structured result.

~/.codex/config.toml 
[mcp_servers.harness-router]
command = "harness-router-mcp"
env_vars = ["OPENROUTER_API_KEY"]
enabled = true
startup_timeout_sec = 10
tool_timeout_sec = 5

# then restart Codex
/mcp

route
route_mcts
03

When one move is not enough, search.

route_mcts explores a supplied side-effect-free state graph. Jev can seed the root once; the remaining simulations are local. The server never executes simulated writes, shell commands, browser mutations, or network mutations. It selects only the first real action.

root
state
quick
0.60
invest
0.78
verify
0.31
inspect
finish
1.00
fallback
4,096 sims · depth 3 benchmark profile
04
Decision latency · 26 Sep 2026

Time to choose a tool.

Two experiments. 24 paired decisions each. Four, eight, or sixteen candidates. Timing stopped at the choice; selected tools were never executed.

Observed decision time

Milliseconds · same scale · lower is faster

route

Ordinary tool choices

9.5× faster · observed mean latency
Codex3,842.1 ms
Harness-router404.5 ms

24/24 choices match CodexAgreement only

route_mcts

Three-step trees · 4,096 sims

12.3× faster · observed mean latency
Codex4,772.4 ms
Harness-router388 ms

24/24 optimal · 100%Codex: 24/24 · 100%

Each experiment has its own Codex baseline and different timing boundaries. Ratios compare observed latency. The MCTS budget was tuned on these same 24 cases.

Methodology & source data

What the clocks include

Codex chose a name only; its timer includes response generation, host scheduling, and dispatch. Router timings include MCP transport and the provider round trip; MCTS also includes local search. Graph preparation is excluded. Pure reasoning time and task completion time were not measured.

How to read the results

These exploratory runs used different workloads in an existing Codex conversation, with cases authored by Codex. They do not isolate the cost of MCTS over ordinary routing. Matching choices measure agreement; the MCTS test checks the highest-reward branch. No Python scripts or OpenAI API calls were used.

route · 24 paired choices

Ordinary tool selection with 4, 8, or 16 candidates. All measured samples are included. Zero router fallbacks or errors. One 695 ms warmup is excluded.

Report ↗ Raw timings ↗

route_mcts · 24 paired trees

4,096 simulations, depth 3, Jev prior enabled. One policy evaluation per response. A separate budget check and this timed rerun each reached 24/24 on the same fixtures; this is not a held-out result. All measured samples are included, with zero final fallbacks or errors. Budget-selection calls and a 412 ms pre-run check are excluded from these timings.

MCTS report ↗ Raw timings ↗ Earlier 32-simulation run ↗
$ 05 · Install

Three steps. Then get out of the way.

Install the latest main branch, expose your OpenRouter key to the local MCP process, and register the server in Codex.

01 · GLOBAL

Install

Keep it isolated and globally available through uv tool.

uv tool install --force \ --with 'mcp>=2,<3' \ 'git+https://github.com/Protocol-Lattice/harness-router.git@main'
02 · ENV

Expose the key

The server can load without it. Jev routing needs it when route is called.

export OPENROUTER_API_KEY="..." command -v harness-router-mcp
03 · CODEX

Register MCP

Add the stdio server, restart Codex, then verify both tools with /mcp.

codex mcp add harness-router -- \ harness-router-mcp codex mcp list
$ 06 · The rule

Route less.
Choose better.

Use the model where reasoning matters. Use Jev where the choice is closed. Use MCTS where the next few moves matter. Skip all of it when the next step is obvious.

View on GitHub Read docs Install MCP