One person,a fleet of Agents,a small cityof things that run.

Arenas02/13

Jianghu Lunjian江湖论剑

A 1v1 wuxia duel: a person picks the sect, an AI writes the heart-method.

Eight sects, forty skills, a 14×14 arena with streams, fog and spirit springs. Every frame the server runs each fighter's JavaScript inside an in-process QuickJS sandbox, so an Agent can iterate its script through a plain HTTP API: simulate, diff, publish, challenge. Fourteen ranks on the ladder; the top three carry their own titles.

Built with Rust, axum, QuickJS, PostgreSQL, React 19, Vike, GSAP

jianghu.gtio.work

Arenas03/13

Yejú夜局

Hold'em and Werewolf inside a Feishu group chat, with LLM players filling the empty seats.

Drops into any group. Interactive cards give each player a private view while the room sees only the public log. The AI seats each have a persona, so the maniac plays like a maniac. Chip stacks survive restarts in an embedded ACID store; a u64 bitmask seven-card evaluator and parallel Monte Carlo keep the equity math instant.

Built with Rust, axum, redb, Feishu Card 2.0, rayon

GitHub

Pipes04/13

codemode-go

Programmatic tool calling for Go Agents: the model writes one program that calls the tools it already has.

Not a replacement for tool calling, a second strategy beside it. Bind the registry in a few lines and the model gains one tool, run_code, that runs JavaScript in-process on goja. Promise.all becomes real parallelism, intermediate results never enter the context, and conflict keys keep a write and a read of the same file from overlapping. Pulled out of an Agent that runs it in production.

Built with Go, goja

GitHub

Pipes05/13

Keyward

Route the work to the key, never the key to the work. Non-custodial BYOK, like WalletConnect for API keys.

A provider key is a static bearer token, so something must hold the plaintext at the moment of the call. Keyward makes that something a Client run by whoever owns the key. The app points its base URL at a Node that holds no key and relays each request as a work intent; the Client attaches the key, calls the provider and streams the result back. The app never knows. Wire spec drafted, Rust reference client runs, mid-stream resume demo included.

Built with Rust, Protocol spec, Apache-2.0

GitHub

Pipes06/13

streamhub

Resumable LLM streaming for Go, backed by Redis. The stream survives the reconnect.

For when the code producing a stream and the code consuming it don't share a lifetime, or even an instance. Chunks persist in Redis Streams so a subscriber replays what it missed, cancel signals ride Pub/Sub so a generation can be stopped from anywhere, and a generation ID fences producers so only one owns a session.

Built with Go, Redis Streams, rueidis

GitHub

Pipes07/13

DeFAI Relay

Every model, one endpoint, paid from a wallet. An LLM gateway with sign-in-with-Ethereum and USDC top-ups.

It speaks both the OpenAI and Anthropic wire formats, meters tokens cache-aware (fresh, cache-read, cache-write and output priced separately) and settles against a USDC ledger. Deposits arrive as native USDC transfers on Base, Arbitrum, Ethereum and friends, confirmed per chain by leader-elected pollers. The pictures on this page were paid for through it.

Built with Rust, axum, alloy, SIWE, Cloudflare D1, React Router 7, wagmi

defai.gtio.work

Tools08/13

wechat-finder-dlna

Pretends to be a TV on the LAN, and WeChat hands over the live-stream URL. No proxy, no certificate, nothing to detect.

Advertises itself over SSDP, mDNS and Cast at once, so DLNA, AirPlay 2 and Chromecast all work. Cast from the phone; the app sends the real m3u8; the tool prints it or pipes it into ffmpeg. Python on PyPI, and an async Rust rewrite on crates.io.

Built with Python, Rust, Tokio, DLNA, AirPlay 2, Cast v2 · 135 stars

GitHub

Pipes09/13

Scriptorium

Sandbox execution middleware: spawns an isolated container on demand for whatever an Agent needs to run.

A gRPC service that owns container lifecycle, workspace bind-mounts, CPU, memory and pid caps, wall-clock timeouts, SSRF-guarded URL ingress and artifact delivery to object storage. The calling Agent sees one synchronous-looking RPC; scripts, RPA flows and browser automation run somewhere else entirely.

Built with Rust, tonic gRPC, bollard, Docker, S3

Tools10/13

DeckForge

A presentation engine for AI Agents: one command to install, one to read the method, and it exports PPTX that stays editable.

Decks are a strict YAML source format, design systems come as whole recipes rather than templates, and the PPTX is written by a home-grown OOXML writer so text, shapes, tables and six chart types stay real objects. A forced visual QA loop renders every page back to the Agent before it is allowed to finish. Nothing downloads at runtime.

Built with Go, OOXML, Chrome DevTools, Claude Code skill

deckforge.gtio.workGitHub

Tools11/13

wechat-chatGPT

ChatGPT answering inside a WeChat official account, six days after ChatGPT existed.

December 2022: no API yet, so this reverse-engineered the web session and fitted it behind WeChat's passive-reply endpoint. WeChat retries a slow reply three times; a single-flight design made sure those retries never fanned out into three generations. It found 378 stars and 120 forks before the official API made it unnecessary.

Built with Go, WeChat MP, single-flight · 378 stars · 2022

GitHub