PROJECTS

Send one prompt to several models at once and read the answers side by side. Responses stream in over SSE, so latency and token throughput show up as they happen.

An LLM-as-judge step scores the outputs, and interactive charts track cost against quality across runs — so you pick a model on evidence instead of vibes.

Type what you want, and the tool drafts a command and shows it in a terminal UI. Nothing runs until you press enter, so you stay in control of the shell.

The context sent to the model is deliberately thin: OS, shell, hostname, working directory, a top-level file listing, and the git branch. File contents, environment variables, command history, and diffs never leave the machine.

Written in Rust with Leptos and compiled to WebAssembly, it talks straight to OpenRouter using OAuth PKCE — no backend to host, and no middleman holding your keys.

Archived, but a clean reference for a fully client-side LLM app.

Runs six workloads — primes, sorting, matrix, hashing, and JSON — in both single- and multi-threaded modes.

Scores are normalized against a reference CPU and exported as JSON, so results stay comparable across different machines.

Periodically deletes messages older than a configurable age threshold, with bulk deletion and adjustable intervals.

Ships as a Docker image or a standalone binary, so it drops into an existing host without extra tooling.

Ingests API call data through webhooks and turns it into real-time cost monitoring, latency percentiles, and filterable trace exploration.

Built with Go and Svelte on top of DuckDB. Archived, but the webhook-to-analytics pipeline still holds up on its own.

Fetches a URL, extracts the readable content, and converts the HTML to Markdown the model can actually use.

Supports pagination, honors robots.txt, and can route through a proxy. Ships as a Docker image.

Proxies transparently between stdio JSON-RPC and a remote MCP server over HTTP/SSE.

Includes full OAuth 2.1 support and compiles to a single ~6 MB static binary — no Node runtime required.

Covers the full OpenRouter API with streaming support, thorough error handling, and extensive configuration options.

A small, dependency-free package for Go services that call LLMs.

Titles and entries are encrypted in the browser or local CLI before they are sent; the server only ever sees ciphertext.

Reach pads through a live web interface, a CLI, MCP tools, or a REST + SSE API, with share links for read-only or collaborative access.

Create custom slugs, set links to expire or work only once, and track clicks.

A single Go binary serves an embedded SvelteKit interface and stores everything in SQLite, with a JSON API for automation.

Decodes Microsoft Safe Links, Proofpoint URL Defense, and similar rewrites locally in the browser, without visiting the destination.

Also available as a JSON API and a Raycast extension for macOS.