Minerva · Skill intelligenceTurn agent sessions into skills.
Read the conversation. See the tool calls. Propose the skill the harness should have loaded — then keep that library in sync.
Skip to content
Named for wisdom — Minerva watches how your agents work, then writes it down as skills.
01Sessions
02Signals
03Skills
The operating problem
Agents repeat themselves. Skills should capture that.
A harness session already contains the workflow: the prompt, the tool calls, the retries, the corrections. Minerva reads those traces and proposes SKILL.md files so the next session does not rediscover the same path.
One honest loop
From a transcript to a skill you can load.
Minerva does not run your agent. It reads what the harness already wrote to disk, extracts deterministic signals, and drafts skills with evidence.
01 / READ
Parse harness sessions
Claude Code, Codex, Cursor, and others — normalized into one Trace model.
02 / SIGNAL
Find the pattern
Retries, corrections, load-gaps, repeated tool sequences, shell families.
03 / PROPOSE
Draft a skill
New, update, or “load this existing skill” — each with session evidence.
04 / SYNC
Keep harnesses aligned
Canonical ~/.agents/skills, linked or copied into writable harness dirs.
Open source · Go CLI + MCP
Install and read your first sessions.
Four commands get you from a binary to ranked skill proposals.
$ go install github.com/abdul-hamid-achik/minerva/cmd/minerva@latest $ minerva init $ minerva sessions --since 7d $ minerva propose --since 30d