NeuroRouter

When AI fails, you keep going

AI sessions break. Corrupted tool chains return 400s. Content filters block legitimate output. Rate limits lock you out mid-task. Runaway agents burn your budget in minutes. And the vendors close the bugs as stale.

NeuroRouter sits between your tools and the API. It blocks secrets, strips wasted tokens, repairs broken requests, and keeps your work alive when everything else fails.

Runs locally. Your API key never leaves your machine except to the provider.

Get The Free Version

The community edition is published at obstalabs/neurorouter. Install it with Homebrew, Scoop, or download the release binaries directly.

brew tap obstalabs/tap
brew install obstalabs/tap/neurorouter

scoop bucket add obstalabs https://github.com/obstalabs/scoop-bucket
scoop install obstalabs/neurorouter

https://github.com/obstalabs/neurorouter/releases/latest

Three Layers

Everything happens before the request leaves your machine. No latency. No infra. No key exposure. No lock-in.

How It Works

Drop-in replacement for your API endpoint.

# Claude Code
ANTHROPIC_BASE_URL=http://localhost:9120 claude

# Codex / OpenAI tools
codex -c 'openai_base_url="http://127.0.0.1:9120"'

# Any tool that supports base URL override
OPENAI_API_BASE=http://localhost:9120 aider

One environment variable. Works with Claude Code, Codex, OpenClaw, Cursor, Aider, Continue.dev, and any tool that lets you set an API base URL. Your API key is forwarded as-is at the HTTP layer — never parsed, never stored, never logged.

# start the proxy
neurorouter --port 9120

# see what would be filtered without sending
neurorouter --port 9120 --dry-run

Concrete Example

A session sends the same file content 3 times, includes a 120KB thinking block, and repeats system instructions in every request. NeuroRouter strips it before the API call. Less wasted context. Lower cost. Same workflow.

Output Purity Score

OPS shows how much unnecessary traffic was removed before billing. A session at OPS 75% means 25% of what would have been sent was noise — caught and removed. Track it in your status line or your logs.

Input Purity Score

ContextSpectre measures IPS — how clean your context stays over time. Without filtering, raw tool output (git diffs, test logs, build output) enters context and gets re-read on every turn. Typical unfiltered sessions score IPS 50-60.

With NeuroRouter Pro active, IPS stays at 90+. The proxy strips noise before it enters context, so the model only sees clean tool results. Less noise in, less re-read tax out.

OPS and IPS are complementary. OPS measures what NeuroRouter removes from each request (output purity). IPS measures what remains in context across turns (input purity). Together they show the full picture — waste prevented at the request layer and waste prevented from accumulating over the session.

IPS metric definition: contextspectre docs/concepts.md

What NeuroRouter Is Not

• Not an observability platform — it transforms traffic, not just logs it
• Not a model gateway — it makes requests better, not just routes them
• Not developer surveillance — patterns, not people
• Not ML-based — deterministic filters, no classifier guesswork

Security

Your API keys never leave your machine. NeuroRouter runs locally, uses environment variables or client passthrough auth, and never stores, transmits, or logs credentials. Keys exist only in process memory for the duration of the upstream request.

This is a structural difference from cloud LLM proxies. A 2026 study found 26 LLM proxy services collecting user credentials. The LiteLLM supply-chain breach (March 2026) compromised thousands of organizations. NeuroRouter eliminates this class of risk entirely — there is no server to breach and no database to leak.

• Zero telemetry, zero analytics, zero phone-home
• No online license validation — Ed25519 signature verified locally
• Verify with lsof -i -P | grep neurorouter — only your upstream
• Build from source: go build ./cmd/neurorouter

Pricing