v1.0 — Claude Code plugin available now

Capture every prompt.
Own the signal.

Promptcellar logs every prompt you send to your AI agent into a structured, append-only file in your repo. Open format, open source, and yours forever — not your vendor's.

Install Promptcellar

$curl -fsSL https://get.promptcellar.io/claude-code | sh

Open a Claude Code session in any git repo to start capturing.Codex coming soon →

How the plugin works Read the PLF spec View on GitHubMIT licensed · zero telemetry · sub-10ms

01 — The pitch

Your code already has a history.
Your prompts don't.

Every commit your AI agent makes started as a prompt — a question, a correction, a “no, do it like this.” Today those prompts vanish into a vendor's database. Tomorrow you can't audit them, replay them, or learn from them. Promptcellar puts them in your repo, next to the code they produced.

Audit, don't guess.

Trace any commit back to the prompt that produced it. Every record has the prompt, the git SHA, the files touched.

Replay, don't redo.

Re-run a session against a different model or a fresh branch. The prompts are the script — the agent is the actor.

Analyze, don't wonder.

Cost, time, who's productive, what works. It's JSONL. jq is enough. Build whatever dashboard you want.

Own, don't rent.

The data is in your repo. Not in a SaaS dashboard you'll lose access to when you switch tools.

02 — How it works · 60 seconds

A file per session.
A line per prompt.

You type a prompt.

…in Claude Code. Same workflow. No new shell, no extra binary on your PATH.

Hooks fire.

Claude Code's native lifecycle hooks (SessionStart → UserPromptSubmit → PostToolUse → Stop) fan into Promptcellar's tiny Go binaries.

A line lands.

One canonical record per prompt, flushed at Stop into .prompts/YYYY/MM/DD/<session>.jsonl.

your-repo/

├── .git/

├── src/

├── README.md

├── ▸.prompts/— append-only log, one file per session

│ └── 2026/05/01/

│ ├── 7b3e9a4f.jsonl← session record shown ↓

│ ├── a2c9e017.jsonl

│ └── f31bd8c2.jsonl

├── ▸.promptcellar/— team capture decision, committed

└── .gitignore

.prompts/2026/05/01/7b3e9a4f.jsonlplf-1

{
  "version": "plf-1",
  "id":          "01HX8K2P7M9NRT3VQYZB4WGFCD",
  "session_id":  "7b3e9a4f-1c84-4d2a-9e21-bd3f7af80c91",
  "timestamp":   "2026-05-01T14:22:31.412Z",
  "author":      { "email": "you@team.com", "name": "You" },
  "tool":        { "name": "claude-code", "version": "1.4.0" },
  "model":       { "provider": "anthropic", "name": "claude-opus-4-7" },
  "prompt":      "Refactor the auth middleware to use the new token store.",
  "git":         { "branch": "feat/auth", "head_commit": "a1b2c3d", "dirty": false },
  "outcome":     { "files_touched": ["server/auth.ts"], "status": "completed" },
  "enrichments": { "input_tokens": 8412, "output_tokens": 1330, "cost_usd": 0.041 }
}

That's the format. It's called PLF. It's an open spec — read it →

03 — The first three things you'll ask

Skeptical? Good.

“What about secrets?”

Built-in redaction. Always on.

Every prompt is run through 222 vendored gitleaks rules covering AWS, GitHub, Stripe, Anthropic, OpenAI, and more — plus a hand-rolled PII layer with Luhn-validated card numbers and ISO-13616 IBAN checks.

.promptcellarignore for team-specific deny rules. Excluded records leave a typed gap so the timeline stays auditable.

“Where does the data live?”

Your repo. That's it.

No SaaS account. No cloud bucket. No telemetry. .prompts/ is a directory you commit alongside your code.

Push it to GitHub, GitLab, your self-hosted Gitea — anywhere git goes, your prompt history goes too.

“Will it slow Claude Code down?”

Sub-10ms cold start. You won't notice.

Hooks are tiny Go binaries bundled inside the plugin. No node, no python, no PATH pollution.

The capture pipeline buffers per session and flushes one canonical record at Stop.

04 — The open standard

PLF is the open format underneath.

Promptcellar is one implementation. The on-disk shape — append-only JSONL, RFC 3339 timestamps, content-addressable session files, schema-versioned records — is a separate open spec called the Promptcellar Logging Format (PLF). Any agentic coding tool can write it. Any tool can read it.

Today: Claude Code. Tomorrow: Codex, Cursor, Aider, your own agent.

Read the spec →View on GitHub

WritersClaude CodeCodex · soonyour agent

→PLF→

ReadersjqNerve Centeryour CIyour dashboard

05 — On the roadmap

What's next.

Coming soon

Codex, coming soon.

Same PLF format. Same .prompts/ folder. Different agent. Email signup for the launch.

In development

Where your prompt signal converges.

A team analytics product over your .prompts/ — live cost and activity dashboards, plus an MCP node that pipes prompts into your workflows. Read-only over PLF.

Join the waitlist →

Capture every prompt.Own the signal.

Your code already has a history.Your prompts don't.