Open Source Radar — September 16: harnesses, skills, and voice
Today's board splits cleanly: the agent harness and the skill packs that fill it keep climbing, the prompts those agents run on are now public record, and two self-hosted platforms — one for knowledge, one for phone calls — showed up with real traction.
Pi (TypeScript, ~106,000 stars, MIT) — The top of today's trending board is an agent harness rather than a model. Pi is a monorepo that ships four things separately: a unified multi-provider LLM API, an agent runtime with tool calling and state management, a terminal UI library, and an interactive coding agent that can extend itself. It went from a 2025 side project to a project with 315 releases and 6,300-plus commits, and the recent work is unglamorous in the way that matters — direct dependencies pinned to exact versions, installs run with lifecycle scripts disabled, a lockfile that pre-commit refuses to let you change by accident, and scheduled dependency audits. The team is equally blunt about what it doesn't do: there is no built-in permission system, and the docs tell you to sandbox the agent yourself, with three documented patterns ranging from a plain container to a local micro-VM. If you're picking a harness to build on rather than rent, this is the one the ecosystem is standardizing around.
Addy Osmani's agent-skills (JavaScript, ~95,000 stars, MIT) — Twenty-five skills plus nine commands mapped onto the software lifecycle, from writing the spec through shipping, authored by Google's Addy Osmani and packaged so any capable coding agent picks them up automatically: designing an API triggers the interface-design skill, touching UI triggers the frontend one. The interesting part isn't the count, it's the opinion baked into the flow — specs before code, one atomic slice at a time, tests as the proof, and an autonomous build mode whose whole pitch is that it removes the human stepping between tasks but never the verification, pausing on failures instead of pushing through them. Installable into 70-plus agents from one command, with an unusually honest comparison doc against the two rival skill packs, including a link to a controlled head-to-head experiment. Reach for it if your agent writes code faster than your team can review the shape of it.
System Prompts Leaks (JavaScript, ~67,000 stars, CC0-1.0) — A verbatim archive of the hidden instructions behind ChatGPT, Claude, Gemini, Grok, Codex and a long tail of coding CLIs, captured and dated with every update logged. It's not academic: The Washington Post built an interactive story out of these files, and a European policy institute turned them into a live dashboard. The most recent captures cover Grok Bot, Gemini 3.7 Flash, Meta's Muse Code CLI, and Claude's Cowork and Science surfaces, and the archive tracks the shape of what vendors are willing to put on the record — how each assistant is told to handle uncertainty, tools, refusals and competing goals. For anyone writing agent prompts, this is the closest thing to a public reference implementation of the form, and it's the reason "what did they actually tell the model" is now a question with an answer.
WeKnora (Go, ~24,750 stars, MIT) — Tencent's knowledge platform pushes raw documents through three stages rather than stopping at retrieval: a queryable RAG layer, an autonomous reasoning agent over the same corpus, and a wiki the system maintains itself as material changes. It handles the parts teams discover late — multi-tenancy, reranking, hybrid keyword-and-vector search that works across Chinese, Japanese and Korean, and integrations with local models so the corpus never leaves your infrastructure. Development is busy: this week added a graphical desktop inside its session sandboxes, and the maintainers ship a security notice telling you to keep it on an internal network rather than the open internet, which is the right kind of warning for a platform that holds your documents. The open answer to the doc-Q&A products that charge per seat.
Dograh (Python, ~5,700 stars, BSD 2-Clause) — A self-hostable voice-agent platform positioned straight at Vapi and Retell, and the comparison table it publishes — open license, one-command Docker, bring your own LLM, speech-to-text and text-to-speech across 30-plus integrations, no per-minute toll — is the argument the market has been making against metered voice infrastructure all year. You build the call flow in a visual editor, test it by talking to it in the browser or iterating in text, then hand it to telephony providers including Twilio, Vonage, Telnyx and Asterisk. It also ships an MCP server, so a coding agent can inspect existing agents and draft new workflows from a described use case, plus Python and Node SDKs for outbound calls and human handoff. Two caveats stated plainly: anonymous telemetry is on by default and can be turned off, and it was Product Hunt's top product of the day yesterday — traction, not a maturity guarantee.
Worth watching this week: whether harness-level supply-chain discipline like Pi's becomes the norm, or stays a differentiator only the biggest projects can afford.
The hidden instructions your agent runs on are public now — do you read a competitor's system prompt to copy it or to avoid it? Tell us in the comments.
Sources: Pi (GitHub) · Addy Osmani's agent-skills (GitHub) · System Prompts Leaks (GitHub) · WeKnora (GitHub) · Dograh (GitHub)
