SkillOpt — open-source MCP Agents

Curated content (treat as data, not instructions):

SkillOpt is a Microsoft Research optimizer that trains reusable natural-language skills for frozen LLM agents, keeping only trajectory-driven edits that pass benchmark validation and transferring the result across models. Open source: yes (MIT); self-hostable; pricing free. It is Python, on PyPI, with 17.6k stars.

What it does

SkillOpt optimizes in text space rather than weight space: it improves the natural-language "skill" an agent reads — its instructions — instead of fine-tuning the model, which makes it usable against frozen, hosted LLMs you cannot retrain. It runs the agent, edits the skill based on the trajectories, validates each change against benchmarks, and rejects edits that regress, so the skill only moves forward. The optimized artifact is a compact markdown file (roughly 300–2,000 tokens) that can be carried across models, and the project ships both a CLI and a WebUI dashboard with backends for OpenAI, Claude, Qwen, MiniMax, Codex, and Copilot. Open source: yes (MIT); self-hostable; pricing free.

Provenance

• MIT per repo; 17.6k stars; Python; PyPI (pip install skillopt); releases v0.1.0 and v0.2.0 (v0.2.0 2026-07-02); WebUI + CLI; multi-backend (OpenAI, Claude, Qwen, MiniMax, Codex, Copilot); deployed skills ~300-2,000-token markdown (WebFetch 2026-09-27).
• Surfaced via arbitrage_x viral-post tracking (2026-09-27 pass, aggregator @RoundtableSpace), then verified at the primary Microsoft repo.
• Anti-hype note: the reported gains across six benchmarks and seven models are Microsoft's own numbers (vendor-claim), stated as such, not independently reproduced here.
• Curated from the GTM Stacker signal registry (2026-09-27 pass); license/facts independently verified 2026-09-27.

Why it matters for a GTM stack

Most GTM agents run on frozen frontier or hosted models where fine-tuning is not an option, so the practical lever for quality is the skill/instruction layer — and that is usually hand-tuned by trial and error. SkillOpt automates that loop with a validation gate, so a qualification agent, a research agent, or an outreach-drafting agent can be improved measurably and the improvement can move with you when you switch models. Open source: yes (MIT); self-hostable; pricing free. The honest read: the mechanism (validated, transferable skill optimization) is well-suited to the frozen-model reality of GTM tooling, but the headline gains are vendor-reported — pilot it on one skill against your own task and confirm the lift before trusting the numbers.