GitHub Repository

An OpenAI-compatible reverse proxy you run yourself. It gives you the features of an AI gateway (guardrails, budgets, rate limits, multi-provider routing) but under your control from your client.

3 starsGo

Personal AI gateway for OpenClaw – tokenomics

Name: Personal AI gateway for OpenClaw – tokenomics
Availability: InStock
Author: crawdog

by crawdog·Mar 3, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidSolve My ProblemBig Brain

OpenAI-compatible proxy with PII masking and token budgets—but LiteLLM, Helicone already do this.

Strengths

•Zero code changes: drops in as reverse proxy for any OpenAI-compatible client; PAT scoping elegant.
•PII masking and jailbreak detection run server-side before API calls, reducing exposure.
•Session logging to disk or Redis with cost attribution stored in git-friendly JSON.

Weaknesses

•Crowded category: Helicone, LiteLLM proxy, Anthropic's own gateway, and enterprise vendors cover this.
•No comparison to alternatives or benchmarks showing why this is better than established solutions.

Post Description

I created a personal AI gateway that runs on your local machine to provide guardrails for any service that communicates with popular LLMs. You can run this as a proxy for openclaw to inject prompts, filter PII, warn if there are jailbreak attempts. Can also work in a distributed fashion.

It creates a personal access token (PAT) based on your API keys from Anthropic, OpenAI or others, then injects policies. It does not store secrets, but instead stores a reference to environment variables storing those values.

Another benefit - if you execute Claude code or Codex using the run command, you can also record session token usage and the prompt traffic. This allows you to store it with your github project as a shared memory.

Run the binary using 'run' for a context that lasts for the command, or use 'start/stop' to initiate the proxy at localhost:8443. There is a web UI you can access to gather stats as well.