▶ Open Beta — Free tier available

Stop paying frontier prices
for mechanical tasks

70% of agentic steps are mechanical — file reads, shell commands, test runs. mesh-router sends them to local models at ~$0. Reserves Claude/GPT for reasoning. Cut your agentic API bill by 70%+.

Get Pro — $15/mo View on GitHub

mesh-sh — routing log

$ mesh-sh execute "add error handling to api/main.py"

# Reasoning Router classifying tasks...

[mechanical] read_file api/main.py → llama-3-8b $0.000

[mechanical] list_dir api/ → llama-3-8b $0.000

[reasoning] plan_changes → claude-3.7 $0.012

[reasoning] generate_patch → claude-3.7 $0.031

[mechanical] write_file api/main.py → llama-3-8b $0.000

[mechanical] run_tests pytest → llama-3-8b $0.000

✓ Done. Cost: $0.043 (saved $0.29 vs all-frontier)

// how it works

A reasoning router inside your agent loop

mesh-router wraps your existing agentic workflow. Each tool call is classified before dispatch. Mechanical tasks never leave your machine.

Incoming task

read_file("main.py")

mechanical

ls, grep, cat, tests

→ Llama 3 / DeepSeek local

reasoning

plan, generate, review

→ Claude / GPT / Gemini

→

⚡

Reasoning Router

Llama-3-70B
local classifier

→

local mesh

Your machines

Ollama · DeepSeek · Llama

frontier

Remote APIs

Claude · GPT · Gemini

Task type	Example	Without mesh-router	With mesh-router	Saved
mechanical	`read 500 files`	$1.50	$0.00	100%
mechanical	`run test suite`	$0.45	$0.00	100%
reasoning	`plan refactor`	$0.08	$0.08	0%
reasoning	`generate code`	$0.12	$0.12	0%

// get started

Up in 5 minutes

Works with any Ollama-compatible local model. No config required to start.

Step 01

Install mesh-sh

pip install mesh-router

Adds the mesh-sh CLI. Wraps your existing agentic tools — Claude Code, Aider, or bare API.

Step 02

Point to your mesh

Configure your local Ollama endpoint and frontier API keys in ~/.mesh/config.toml. Tailscale mesh? Set your worker IPs.

Step 03

Run your loop

Replace claude with mesh-sh. The router handles classification automatically. Watch costs drop in real time.

Step 04 (Pro)

Cloud Router

Pro tier unlocks the cloud-hosted Reasoning Router — a 70B model tuned on task-type classification. Optimal distribution across your whole fleet.

// pricing

Start free, scale as you grow

Bring your own hardware for Open Source. Pro adds cloud routing. Managed handles everything.

Open Source

Self-hosted local workers only
Ollama + DeepSeek + Llama support
Basic mechanical/reasoning classifier
Works with Claude Code, Aider, bare API
No cloud router
No managed GPU workers

View on GitHub

Stop paying frontier pricesfor mechanical tasks

A reasoning router inside your agent loop

Up in 5 minutes

Start free, scale as you grow

Stop paying frontier prices
for mechanical tasks