agsearch is currently something you run beside your coding agent. As a skill it becomes something the agent runs itself.
The difference in practice: instead of the agent having no recollection of an earlier conversation, it searches the transcripts already on disk and answers from them.
"what did we decide about the webhook retry backoff?"
→ agent runs agsearch -n "webhook retry backoff", reads the ranked hits, answers
Why this is cheap
agsearch -n "query" already exists and prints ranked plain text. That is the whole integration surface.- A Claude Code skill is a markdown file. No server, no protocol, no new dependency, and nothing added to the CLI.
Scope
- A
SKILL.md describing when to reach for agsearch and how to read -n output.
- Guidance on query construction.
-n is exact substring, AND across terms, so passing a whole question returns nothing. The skill needs to teach keyword extraction, and that is most of the actual work here.
- A note on volume:
-n caps its output and reports ... and N more. Narrowing the query beats paginating.
Not this issue
An MCP server is the heavier version of the same idea. The skill gets most of the value for a fraction of the effort, and it is worth learning whether the skill is used before building a protocol around it.
Not started. Contributions welcome, and if you would use this, say so here.
agsearch is currently something you run beside your coding agent. As a skill it becomes something the agent runs itself.
The difference in practice: instead of the agent having no recollection of an earlier conversation, it searches the transcripts already on disk and answers from them.
Why this is cheap
agsearch -n "query"already exists and prints ranked plain text. That is the whole integration surface.Scope
SKILL.mddescribing when to reach for agsearch and how to read-noutput.-nis exact substring, AND across terms, so passing a whole question returns nothing. The skill needs to teach keyword extraction, and that is most of the actual work here.-ncaps its output and reports... and N more. Narrowing the query beats paginating.Not this issue
An MCP server is the heavier version of the same idea. The skill gets most of the value for a fraction of the effort, and it is worth learning whether the skill is used before building a protocol around it.
Not started. Contributions welcome, and if you would use this, say so here.