The plugin · v1.0

Promptcellar for Claude Code.

A Claude Code plugin that captures every prompt you send to the agent and stores it in your repo as structured PLF.

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 →

01 — What gets captured

Every prompt. Every outcome.

Secrets and PII are stripped before anything touches disk — redaction is on by default. Everything else is captured verbatim.

Verbatim prompt text + git author identity.
post-redaction · 222 gitleaks rules + PII layer
Git state at prompt time: branch, HEAD, dirty/clean.
Outcome: files touched, commits made, status.
completed · interrupted · errored
Enrichments: input/output tokens, estimated cost, wall-clock duration.
Tool & model versions, for reproducibility.

.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 }
}

02 — In-session controls

Everything is a slash command.

No binary on your PATH. No new shell. The plugin's only surface is /promptcellar:* inside Claude Code.

Promptcellar capturing a prompt inside Claude Code: slash command runs, prompt is captured, JSONL record lands in .prompts/ — Plugin in action — install, type a prompt, watch the record land.

Command	What it does
/promptcellar:status	Resolved config + capture counts + .prompts/ location
/promptcellar:log [N]	The last N captured prompts in this repo
/promptcellar:doctor	Diagnose hook wiring, transcript parser, perms
/promptcellar:enable / :disable	Per-repo (committed) or --for-me / --global
/promptcellar:version	Plugin + transcript-adapter version
/promptcellar:uninstall	Remove plugin entry; data left intact

03 — Three-layer config

Repo decides. You opt out. Machine kills.

Resolution order — most-restrictive wins.

01 · Global

~/.promptcellar/config.json

“Stop capturing on this whole laptop today.”

02 · Repo · committed

.promptcellar/config.json

The team's decision. Travels with the code.

03 · Personal · gitignored

.promptcellar/config.local.json

“The repo captures, but skip mine.”

04 — Privacy & redaction

The redaction layer.

Four passes, in order. The first match wins.

222 gitleaks rules

AWS, GitHub, Stripe, Anthropic, OpenAI, Slack, Datadog, GCP, Cloudflare — vendored at build time.

Built-in PII layer

Luhn-validated card numbers. ISO-13616 (MOD-97) IBAN checks. Email and phone fall through to context-aware classifiers.

.promptcellarignore

Team-specific deny rules. Authoritative — overrides everything else, including allows.

.promptcellarallow

Narrow exceptions for false positives. Last in the chain. Local-first.

Matches never touch disk. Excluded records leave a typed gap so the timeline stays auditable — you can see {"status":"excluded","reason":"secret"} in the JSONL.

05 — How capture works

Four hooks. Six binaries. Sub-10ms.

Hook	What we do with it
SessionStart	Registers a session id, writes header line, picks .prompts/<date>/<id>.jsonl path.
UserPromptSubmit	Captures the verbatim prompt + git author + branch HEAD; runs through the redactor.
PostToolUse	Buffers per-tool outcomes (files touched, exit codes) into the open record.
Stop	Flushes a single canonical PLF line. Append-only. fsync optional.

Hooks buffer per session and flush one canonical record at Stop. No partial lines. No lock contention across sessions because each session writes to its own content-addressable file. See IMPLEMENTATION_PLAN.md for the full story.

06 — Reading your data

It's just JSONL. jq is enough.

Last 10 prompts across all branches

cat .prompts/**/*.jsonl \
  | jq -r '.prompt' \
  | tail -n 10

Total cost on a branch

jq -s '[.[] | select(.git.branch=="feat/auth")
  | .enrichments.cost_usd] | add' \
  .prompts/**/*.jsonl

Prompts that touched a path

jq -r 'select(.outcome.files_touched[]
  | contains("server/auth"))
  | .prompt' .prompts/**/*.jsonl

Browse all examples on GitHub →