Prudence

Prudence is a local-first, open-source growth coach for developers who build with AI coding agents. It records how you actually worked with your agent and links that record to what became of the code in git, so you get the feedback a colleague or a code review would normally give you.

Status

0.3.0: it captures your sessions, links them to your commits, follows what became of the code and where the tokens went, and writes a review of it: every number computed from your own record, with an optional model-written paragraph on top that may only quote those numbers. A native macOS menu-bar app (apps/mac) shows the same record with charts. prudence ask answers a question from the record.

Principles we build on

• The record is yours and stays on your machine.
• Capture is opt-in per repository, with a metadata-only mode.
• Secrets never leave the store.
• Findings, not scores.
• Sources on everything.

Platforms

macOS and Linux. Windows is untested.

Install

Prudence needs Python 3.12 or later and uv. If you don't have Python yet, uv can install it for you.

uv tool install --python 3.12 prudence-dev

Optional extras:

• prudence-dev[mcp] adds the prudence mcp server used by the Claude Code plugin.
• prudence-dev[model] adds the Anthropic SDK, for prudence review --explain and the prose half of prudence ask.

uv tool install --python 3.12 "prudence-dev[mcp,model]"

The menu bar is a native macOS app now rather than a Python extra; see apps/mac/README.md.

Quick start

prudence init --scan
prudence init --enable <repo> --level full
prudence ingest
prudence sessions --last 30d
prudence outcomes --project <repo>
prudence usage --last 30d
prudence observations --project <repo>
prudence show --session <id>
prudence facts --last 30d
prudence review --project <repo>
prudence review --explain
prudence ask "what did I spend tokens on last week"
prudence hooks install
prudence export

init --scan lists the repositories found in your agent history without changing anything. init --enable turns recording on for one repository at a time, at full capture or metadata-only (use --level metadata-only for a repository whose content you would rather not store). ingest reads what the agent recorded since the last run. sessions lists what happened; outcomes follows the commits each session produced forward, to what survived; usage follows the tokens and active time back, to what they were for; observations joins the two, inside one project, when your own data supports it. show prints one session's full record; facts prints the behaviour counts every session was scored on. hooks install adds the git-state hooks Prudence uses to catch what happens around each turn; it edits Claude Code's settings file, shows the diff first, and prudence hooks uninstall reverses it. review stores a review of the period since the last one (or --last 14d, --month 2026-09) and prints it; it refuses to write an empty one until enough new sessions and matured commits exist (--force overrides). --explain adds a model-written paragraph, only when a model is configured (prudence config model); the paragraph is checked against the review's own numbers and refused if it invents one or grades you. ask retrieves the sessions the question is about and, unless you pass --no-model, asks the model to answer from that evidence alone. Before every model call Prudence prints what it is about to send.

The first completed ingest ends with a "first look": three to five facts about the history it just read, each with the command that shows more.

What it can tell you today

• prudence outcomes --project <repo>: "62% of the lines from your last 90 days are still in the tree at head, with 78% coverage (41 fact-attributed commits, 12 inferred)."
• prudence usage --last 30d: "development 55%, research 25%, conversation 15%, debugging 5% of this week's tokens."
• An observation at the end of outcomes: "In one project, your sessions that ran a formatter still had more lines at head than the ones that did not (88% vs 74%, 14 sessions each side)."

Claude Code plugin

Prudence also ships a Claude Code plugin that brings your recorded history into a session: /prudence:sessions, /prudence:outcomes, /prudence:usage, /prudence:recall, /prudence:review and /prudence:ask skills, and an MCP server an agent can query directly (ask there returns the evidence only, because the agent calling it is already a model). See plugin/README.md. To try it from this repository without installing it:

claude --plugin-dir ./plugin

What it records and what it never records

Raw session data (transcripts, tool results, file-history snapshots) is archived unmodified, byte for byte, in a local store with owner-only file permissions (mode 0600). Derived tables built from that archive hold counts, classes, and keyed line hashes, never message text: no prompt or response text is written to any derived table, at any capture level.

metadata-only withholds more: file paths and line hashes are not stored, so that level records shape (how much changed, when, in which class of command) without a fingerprint of the content itself.

Nothing Prudence records is uploaded anywhere. If you plan to enable it on a work repository, check your employer's policy first. Observations, the correlations between your own behaviour and your own outcomes, are computed the same way: they never leave this machine, and they never compare you with anyone else, only with yourself.

Contributing

See CONTRIBUTING.md.

License

Apache-2.0. See LICENSE.