Use with AI agents
Increasingly, the thing integrating AgentSight is not a developer reading this site — it is a coding agent working in your repository. These docs are built for that reader too, and the SDK repository ships three packaged skills that cover the whole lifecycle: getting integrated correctly on the first pass, bringing your existing history over, and working with the integration every day after.
The three skills
Both live in the repository in the emerging agent-skills convention — a SKILL.md entry point plus reference files — and they install together:
skills/agentsight-integration is the integration workflow, for a codebase that does not yet record to AgentSight (or a new service being brought online). An agent following it will:
- Read your code first and ask only what code cannot answer — which of the ids it found is your durable conversation id, which tool functions you consent to record, which deployment is production. Every question arrives with what was already determined, and every optional question has a default, so silence still produces a working integration with its gaps named.
- Write the answers down in an
AGENTSIGHT.mdat your repo root, so a later run continues instead of re-interrogating. - Verify before claiming success: run your app with
AGENTSIGHT_FILE_EXPORTERset and read back the spans it recorded — checking turn durations, message text, tool names and token counts against what the integration promised. No key needed for that loop, and the dump doubles as a byte-level answer to "what would leave my process?"
skills/agentsight-migration covers the conversations that happened before you integrated — in a legacy database, a previous chat platform, or a vendor export. Point an agent at that source and it inspects it read-only, interviews you about the mapping (which field is the conversation id, which timezone the timestamps are in, what happens to every row that cannot be represented), and produces a JSON file it validates with a script the skill ships. You upload the file yourself on the dashboard's migrate page — the import routes are deliberately not on the API-key plane, so the agent builds the file and never sends it. See Importing existing history.
skills/agentsight is the knowledge base, for everything after: the full tracking and API surface, the judgment calls the docs teach a human — when the quickstart shortcut fits and when the explicit form is required, why wrap() is not optional on a streaming handler, where shutdown() must be wired, which tools a framework already reports — plus what powers every dashboard metric, the embeddable dashboard, and a symptom-to-cause table for debugging an empty chart. Day-to-day work in an integrated repo loads only this skill, not the integration workflow.
The split is deliberate, and it runs along one axis: the tense and location of the data. Conversations that have not happened yet are the integration skill's; conversations that happened somewhere else are the migration skill's; conversations already in AgentSight are the knowledge skill's. An interview and verification protocol matters enormously once and is dead weight in every session after, so each skill tells the agent when another one applies, and both workflow skills link into the knowledge skill's reference files — which is why they should be installed together.
Installing them
With the skills CLI — works with Claude Code, Cursor, Codex, and most other coding agents; it will offer all three, install all three:
npx skills add agentsightio/agentsightioIn Claude Code, the repository is also a plugin marketplace, which gets you all three skills plus update tracking through the plugin manager:
/plugin marketplace add agentsightio/agentsightio
/plugin install agentsight@agentsightEither way, then ask in plain words — "add AgentSight to this service" runs the integration; "can I get my old conversations in?" runs the migration; later, "why is the escalation chart empty?" or "record feedback from our UI" draws on the knowledge skill. Any harness that can read files can use the skills without an installer too: copy all three folders (skills/agentsight-integration, skills/agentsight-migration and skills/agentsight) from the repository side by side into your project and point the agent at a SKILL.md — both workflow skills link into the base skill's reference files, so either one without it is incomplete.
If you'd rather feed it the docs
Every page on this site has a Copy as Markdown button and serves its raw markdown, so pasting a page into a conversation loses nothing. The pages an integrating agent gets the most from, in order: Quickstart, Turns & Messages, Streaming, Deployment & limitations, and What the SDK sends for the privacy conversation.