Skip to content
Kapacitor docs

What is Kapacitor

Pair programming used to mean two engineers sharing context as they worked. Now it’s an engineer and an agent — and the agent’s reasoning disappears the moment the chat scrolls. The PR ships; the why doesn’t. Reviewers see a diff with no co-author to ask, teammates inherit code nobody can explain, and future-you reopens the file in six months with a fresh mystery.

Kapacitor isn’t observability for your agents, and it isn’t a transcript archive. It restores the shared context that disappeared when pair-programming turned solo — so a reviewer, a teammate, or future-you can ask “why did we pick this?” and get a grounded answer from what actually happened.

Concretely: it records every coding-agent session your team runs — Claude Code, Codex, Cursor — and surfaces them in a real-time dashboard, with full transcripts queryable from the CLI, the web UI, and (via MCP) from inside the agent itself.

This page is the lens for everything else in the docs. If you’re trying to decide whether Kapacitor is right for you, start at kapacitor.ai instead.

The two pieces

Section titled “The two pieces”

Kapacitor has exactly two moving parts you interact with:

  • Your tenant — a Kurrent-hosted Kapacitor instance at https://<your-github-org>.kapacitor.ai. It stores your events, projects them into the dashboard, and serves the web UI. Kurrent provisions and operates it; you never deploy server infrastructure yourself.
  • The CLI — npm install -g @kurrent/kapacitor. It installs hooks into Claude Code and Codex, streams transcripts to your tenant in real time, and provides the commands you’ll use day to day (kapacitor recap, kapacitor review, kapacitor eval, …).

The CLI is open source. The server is not.

What gets recorded

Section titled “What gets recorded”

Once the CLI is set up, every coding-agent session is captured automatically:

  • Session lifecycle — start, end, interruptions, context compaction
  • Full transcripts — streamed to your tenant with sub-second latency, so a teammate watching the dashboard sees turns appear as the agent emits them
  • Subagent trees — every spawned subagent with its own transcript and its own token tally
  • Tool calls — every Bash, Read, Edit, Write, etc. with timing, results, and the exact diff applied
  • Thinking blocks — the agent’s reasoning, captured alongside the turn that produced it
  • Token consumption — input, output, cache reads, and cache writes, per turn and per subagent
  • Repository context — git repo, branch, and PR linkage

You don’t run separate “record” commands; the hooks fire automatically once they’re installed.

What you do with it

Section titled “What you do with it”

Four primary surfaces:

  • The dashboard — https://<your-org>.kapacitor.ai. Browse repos, sessions, agents. Watch sessions live as agents run. See ownership and sharing.
  • Recap, recall, and review — kapacitor recap, kapacitor errors, kapacitor validate-plan, the auto-installed kapacitor-sessions MCP server, and kapacitor review <pr-url>. Replay a session when you know which one, let your agent search past sessions when you don’t (“have we worked on this before?”), and get PR review grounded in the implementation transcripts. Run from your terminal, or invoke as skills/MCP tools inside the agent.
  • Evaluations — score every session against safety, plan adherence, quality, and efficiency using LLM-as-judge. Findings flow back into per-repo clusters that admins curate and re-inject at the top of the next session, automatically. One session’s mistakes become the next session’s guardrails. The most distinctive thing Kapacitor does — see Evaluations.
  • Hosted agents — launch Claude Code or Codex in an isolated worktree directly from the dashboard, controlled by a daemon process running on your machine. Multiple teammates can drive the same hosted session.

Where to go next

Section titled “Where to go next”