Prompt History

Every prompt you have ever typed into Claude Code and Codex is already sitting on your disk. This reads it back.

pipx install prompt-history && prompt-history

Your browser opens. There is everything you have asked an AI to build, searchable, grouped, one click from your clipboard.

No dependencies. No account. No network. It never writes to your session files, it only reads them.

---

Why you want this

You spent six months writing prompts. Some of them were really good. You have no idea where any of them are.

They are in ~/.claude/projects and ~/.codex/sessions, buried in JSONL transcripts next to a few hundred megabytes of tool output and model replies. Technically readable. Practically gone.

This pulls out your half of the conversation and nothing else.

• The prompt you wrote at 2am that finally fixed the build
• That one perfect refactor instruction you want to reuse
• Six months of how you actually talk to these tools, as a file you own

---

Get it

pipx install prompt-history     # recommended
uv tool install prompt-history  # or this
pip install prompt-history      # or the classic

Python 3.9 or newer. That is the entire dependency list.

---

Use it

The browser UI

prompt-history

Search across everything, browse by tool then project then session, copy any prompt with one click, download the lot as a zip.

It binds to 127.0.0.1 and serves a single HTML file. Nothing is uploaded, because there is nothing to upload to.

The terminal, if you live there

prompt-history list                         # everything, newest first
prompt-history search postgres migration    # both words, anywhere
prompt-history stats                        # what you have, at a glance
prompt-history sessions                     # every session, newest first
prompt-history sources                      # where it looked and what it found

Some real ones:

# Your long prompts from this quarter, as Markdown
prompt-history export -f md --since 2026-07-01 --min-words 60 -o good-prompts.md

# Everything you ever said to Codex about auth
prompt-history search auth --tool codex --full

# One project, one archive
prompt-history export -f zip --project checkout-api -o checkout-prompts.zip

# Pipe it anywhere
prompt-history list --json | jq -r '.[] | select(.words > 100) | .text'

--tool, --since, --until, --min-words, --max-words, --project, --exclude-project, --search and --no-slash-commands work on every command, including serve. Filter once, and the UI, the exports and the zip all agree.

The zip, organised

prompt-history export -f zip

prompt-history-2026-09-21/
├── README.txt                  what is in here, and the counts
├── index.json                  the full machine readable snapshot
├── prompts.csv                 one row per prompt, opens in any spreadsheet
├── prompts.md                  everything, one document
├── prompts.txt                 everything, no markup
├── by-session/
│   └── claude/checkout-api/
│       └── 2026-09-14-fix-the-stripe-webhook.md
└── by-date/
    └── 2026-09-14.md

by-session is the folder you actually read. Each file is one conversation, your prompts in the order you typed them, with the project, model and branch at the top.

As a library

from prompthistory import PromptHistory

ph = PromptHistory(since="2026-01-01", tool="claude")

print(len(ph.prompts()))

for hit in ph.search("rate limit"):
    print(hit.timestamp, hit.project, hit.text[:80])

ph.write("q3.zip")                 # format inferred from the suffix
markdown = ph.render("md")         # or get it as a string

Every CLI flag is a keyword argument, and a typo raises instead of silently returning everything.

Prompt carries text, timestamp, project, session_id, model, git_branch, turn, words and kind. Session carries the prompts plus where they came from.

---

Configure it

prompt-history config --init     # writes a commented template
prompt-history config --show     # prints the values actually in effect

Lives at ~/.config/prompt-history/config.toml, or config.json if your Python is older than 3.11.

[sources]
claude_code = true
codex = true
include_subagents = false   # sub-agent prompts are written by the model, not you

[filter]
min_words = 5               # stop "yes", "continue" and "go on" from filling the UI
skip_slash_commands = true

[server]
port = 7777
open_browser = true

Config file, then PROMPT_HISTORY_* environment variables, then CLI flags. Last one wins.

---

Where the prompts come from

Source	Path	What it gives
Claude Code transcripts	~/.claude/projects/<project>/<session>.jsonl	Full text, timestamps, project, model, branch
Claude Code recent history	~/.claude.json	Prompts whose transcript is gone
Codex rollouts	~/.codex/sessions/**/rollout-*.jsonl	Full text and metadata
Codex composer history	~/.codex/history.jsonl	Everything typed at the composer
Codex thread index	~/.codex/state*.sqlite	Opening prompt of threads with no rollout file

CLAUDE_CONFIG_DIR and CODEX_HOME are respected, lists included. Windows %APPDATA% too.

Three Codex rollout formats are supported, current and legacy, because the on-disk shape has changed a few times.

Thrown away: tool results, tool calls, assistant turns, compaction summaries, [Request interrupted] markers, sub-agent prompts, and injected context like <system-reminder>, <environment_context> and <user_instructions>. What is left is what you typed.

When a history log repeats a prompt a transcript already has, you see it once. stats tells you how many were folded together.

---

What it cannot do

Claude Desktop and ChatGPT Desktop chats are not files on your disk. They are in encrypted app storage or only on a server.

Codex makes this visible: if you use the desktop app, the thread titles are cached locally but the prompt text is not. Prompt History counts those threads and tells you, rather than quietly pretending they are not there.

If somebody tells you they can export your Claude Desktop history from local files, check what they are actually reading.

---

Privacy

Your prompts contain your API keys, your client names and your worst code. So:

• Session files are opened read only and never modified
• The server binds to loopback, and there is no outbound request anywhere in the codebase
• No telemetry, no analytics, no update check
• Zero third-party packages, so there is no supply chain to trust but this one

It is roughly 2,300 lines of stdlib Python plus one HTML file. Read it yourself, it will not take long.

---

Contributing

Adding a source is one file.

prompthistory/
├── api.py             PromptHistory: scan, filter, snapshot
├── config.py          file, env and flags, layered
├── model.py           Prompt, Session, text cleaning, timestamps
├── export.py          md, json, csv, txt, zip
├── server.py          loopback HTTP
├── cli.py             the commands
├── sources/
│   ├── claude_code.py
│   └── codex.py       <- copy this one
└── web/index.html     the whole frontend

Write a module in sources/ that returns list[Session], call it from PromptHistory.refresh(), add a fixture to tests/conftest.py. Done.

git clone https://github.com/alshell7/prompt-history
cd prompt-history
pip install -e ".[dev]"
pytest

Cursor, Aider, Gemini CLI and Zed all keep local transcripts. Pull requests very welcome.

---

MIT. Icons from Tabler, MIT.