Agent runtime, not raw API
skillterm invokes a headless agent runtime — Claude Code by default, Codex as an alternative — so the generator gets WebSearch, WebFetch, Bash, and file I/O for free. The CLI orchestrates; the agent reasons.
← home · features
skillterm is the orchestrator; a headless agent runtime does the reasoning. Together they turn documentation into durable, Tab-aware skills. Here is the whole surface.
2
agent runtimes (Claude Code, Codex)
3
shells: zsh, bash, fish
3
generation sources
MIT
licensed, in Rust
skillterm invokes a headless agent runtime — Claude Code by default, Codex as an alternative — so the generator gets WebSearch, WebFetch, Bash, and file I/O for free. The CLI orchestrates; the agent reasons.
Every skill is a SKILL.md with YAML frontmatter and a markdown body, following the public Claude Agent Skills convention. The same file an agent reasons over is documentation a human can read.
eval "$(skillterm init zsh)" wires skillterm into zsh, bash, or fish. On each Tab, the hook calls skillterm complete, which serves completions from the matching SKILL.md.
skill-creator (the format spec and quality checklist) and saas-detector (tool classification and docs discovery) load as context on every generation. Improving skillterm means improving these skills.
Generate from a system man page (default), a specific documentation URL (--from-url), or an open web search (--from-search). Dry-run, force overwrite, and runtime override are all flags.
Generation is expensive and happens once per tool. Tab-time lookup is cheap and happens forever. The durable artefact — a SKILL.md on disk — is what makes that split work.
Skills are managed like packages: search local and remote, install from a registry or gh:user/repo, list installed, and publish your own with --dry-run validation.
skillterm searches three directories in order, so a repo can pin per-project overrides that win over your personal and system skills without clobbering anything global.
Turn allow_web_search and allow_web_fetch off in config.toml and the agent still has Bash and local file access — enough to build a skill from a system man page and --help output.
Go deeper
MIT-licensed, single binary, SKILL.md as the durable artefact.