Quickstart

Install paperctl, capture your first agent session, then prove the value — inspect it, price it, and find it by search.

In the next ten minutes you’ll capture a real AI coding session, open a complete record of everything your agent did, and find it again just by describing it. That session is the first entry in something bigger: a permanent, searchable history of your team’s AI work.

Before you start

  • A paper console account
  • Claude Code, Codex, or Pi installed and working in your terminal
  • A real project where an agent session is safe to run

Milestone 1 — prove capture

paperctl records the session locally while you work — the conversation, tool activity, token usage, and cost. How capture works explains exactly what ends up in a session, worth a skim before your first run.

Install paperctl

curl -fsSL https://download.papercompute.com/install | sh

No sudo — the installer puts paperctl in ~/.local/bin and adds it to your PATH via your shell rc. Open a new terminal if the command isn’t found immediately, then confirm:

paperctl --version

Expected: the version number, for example paperctl 0.21.1.

Authenticate

paperctl login

Follow the browser flow, then confirm who you’re signed in as:

paperctl whoami

paperctl status shows the rest: the daemon running, auth healthy, and your org and gateway selected. You’re wired up — everything you capture from here lands in your console.

Run a captured session

From a project directory, launch your agent through paper:

paperctl start claude

Then use the agent normally. (Codex authenticates to OpenAI a few different ways — pick the path that matches you.)

For a first session, give it a small but real investigation rather than a summary request — real work produces searches, file reads, and decisions worth capturing:

Find where authentication errors are handled in this project, explain the current behavior, and suggest one small improvement. Don’t change any code until you’ve inspected the relevant files.

Confirm the session was captured

paperctl sessions list

Your session is the top row — name, model, status, turn count, cost — and it’s already in paper console. That’s milestone one: work that used to vanish when the terminal closed is now permanent history.

Milestone 2 — prove the value

Captured is only half the point. paper console is where the session becomes history you can explore — open it, price it, and find it again by meaning.

Open the session

Click the session in paper console. Everything your agent did is in front of you as one explorable timeline: Chat is the conversation you remember having; flip to All activity and see what was happening underneath it — every tool call next to its result, subagents nested where they ran, the decisions between your prompt and the answer. Session Info prices the whole thing: cost, duration, tokens. “What did it actually do?” is now a click, not a reconstruction.

Explore a session detail pageLive demo — the activity timeline and the Session Info rail, filled with sample data
Set up Grafana dashboards sub-agent
Completed4h agoJason Carter
Token is in 1password as grafana-sa in the infra vault. Wire it up and verify the dashboards tab works.
Edit.mcp.json
Bashop run -- npx mcp-grafana --check
Wired and verified. .mcp.json now references op://infra/grafana-sa/credential, and the launcher resolves it at start — no secret ever lands in the repo.
The MCP handshake succeeds and the dashboards tab lists all three generated dashboards.
… 15 more turns in the full session
Live demo — switch to All activity and toggle the kind chips; click a turn to expand it
Live demo — hover or tap the turn histogram; expand Session details

Understand session details tours the whole page.

No session name to remember, no branch, no commit, no scrolling through terminal history — just describe what happened:

paperctl search "improvement to auth error handling"

Search matches meaning, so the words don’t have to be the ones the agent used. Results come back most-relevant first, each naming its session so you can jump straight back in. A session becomes searchable within about a minute of capture, so if it’s not there yet, that’s the reason. From now on, everything you capture is findable this way — Search session history has more patterns worth stealing.

Make capture the default

You typed paperctl start claude this time. One line in your shell rc means you never have to again: shell routing makes plain claude, codex, and pi route through paper on their own. From that point forward, every session is captured automatically — the history builds itself.

What compounds from here

Today it’s one session. In a few weeks it’s dozens — and that’s when the history starts answering questions you couldn’t ask before: how was that dependency problem solved, what has anyone tried on this service, which workflows burn money without finishing. Insights surfaces those patterns as real usage accumulates, and the sessions worth repeating become Skills — the best work anyone on your team has done with an agent, packaged so everyone’s agent can do it. The session you just captured is entry one.

If the session doesn’t appear

The most common cause: the agent was started directly (plain claude) instead of through paperctl start. Launch it through paper — or set up shell routing so that can’t happen — and see Troubleshooting for the full checklist.

Frequently asked questions

Do I need to set up a gateway or backend before I start?+
No. When you create an organization, paper console provisions a default gateway and backends, and the paperctl CLI auto-discovers them on login. You only configure backends to customize routing or set a shared team key.
Does paper change how I use Claude Code?+
No. Start the session with paperctl start claude, then use Claude Code normally. paper captures the session in the background.
Why don't I see my session in paper console?+
The most common cause is starting the agent directly instead of through paperctl start. See Troubleshooting for the full checklist.
Can I capture a client other than Claude Code?+
Yes. Codex (paperctl start codex) and Pi (paperctl start pi) are supported the same way, and tools like Conductor can route through paper too.
Copied to clipboard