Imprimatur

See every doc edit your AI agent makes as tracked changes, until you accept it.

Imprimatur is a VS Code extension and a Claude Code plugin (hooks and an MCP server). Imprimatur is Latin for "let it be printed": the approval stamp.

• Review what the agent wrote. Its edits stay marked in the editor and the Markdown preview until you accept them. Committing does not clear them.
• Never miss what it asked you. Questions, permission requests and "test this" steps from every session land in one list that keeps itself up to date.
• Know where you left off. Tasks, decisions and design documents live in one local database that the agent reads and writes through MCP.

Tracked changes until you accept

Added lines are green. A changed word sits next to the old one in a red box. A deleted block leaves a marker. ✓ Accept a block, a file or a whole agent edit, and the view moves to the next change. The Agent Change Graph lists every agent edit in the repo, one lane per Claude session, each with a one-sentence description.

Waiting on you

The agent's questions, permission requests and "test this" steps, across sessions, each written to make sense on its own and tied to its task. After each turn a small model ticks what you answered or what was done. Scan history finds asks from before you installed Imprimatur.

Where you left off: tasks and 👉

Each task keeps its to-dos, questions, decisions, notes, ADRs, PDRs and tests. The 👉 pointer marks where work stopped. The agent gets it at session start; you see it on Home, in the Tasks tab and in Claude Code's status line.

Living design documents: Architecture, Product, Glossary

Every repo has three documents kept in sections: Architecture, Product and Glossary. The agent updates a section when it records a decision and links the two. They stay in step with docs/ARCHITECTURE.md, docs/PRODUCT.md and docs/GLOSSARY.md.

Analyses kept, and /study

Every analysis, plan or research note the agent writes for a task is kept on the task, one version per change. /study [#92](https://github.com/halilural/imprimatur/issues/92) writes a study note on a topic, then quizzes you and keeps a mastery score, so the next study asks the weak questions first.

Agent setup audit and recommendations

Agent setup shows every file that shapes a coding agent, in the repo and your home folder: Claude Code, Cursor, Copilot, AGENTS.md, Gemini, Windsurf, Cline, Aider. Recommendations lints them with agnix and, on request, reviews the project against current best practice. Each finding stays as a record: open, applied or dismissed.

More

• Issue tracker sync: task status follows GitHub Issues or Jira, both ways.
• Process checks (opt-in per repo): block --no-verify, keep code off main, warn when code starts before its ADR.
• Unread marks: edits and document sections the agent changed that you have not read.
• Edit history: the agent's edits to a file, newest first, each with its diff.
• Mermaid diff (off by default): changed flowcharts colored in the preview.
• Archive: edits and asks older than 7 days leave review on their own.
• Plugins: optional local plugins hear every change and can add commands.
• Other agents: Cursor and Codex get the same MCP tools.

Install

Imprimatur has two parts. Install both.

1. The Claude Code plugin (hooks and the MCP server):

claude plugin marketplace add halilural/imprimatur
claude plugin install imprimatur@imprimatur

2. The VS Code extension, from the Visual Studio Marketplace or Open VSX:

ext install halilural.imprimatur

Or install a .vsix from the GitHub releases (Extensions view → Install from VSIX).

To install from a clone instead, with the status line, see Setup options.

Requirements

• Node.js 22.13 or later, and the claude CLI on your PATH (the model features use it).
• VS Code 1.100 or later.
• Linux, macOS or WSL. On WSL, keep the database on the Linux side, not under /mnt/<drive>: Imprimatur refuses a database there.

Quick start

1. Install both parts, then reload VS Code.
2. Ask Claude to edit a Markdown file in a repo.
3. Open the file's preview: the changes are marked. Click ✓ Accept on a block.
4. Click Agent Graph in the status bar for every edit and what waits on you.
5. Open the Imprimatur icon in the activity bar for Home, Tasks and Analysis.
6. Markdown is marked in the preview only; set imprimatur.showIn to both for editor marks too.

How it works

1. Before each agent edit of a Markdown file, a hook copies the file to .claude/imprimatur/baseline/ and logs the edit.
2. The extension diffs each open file against its copy. Accept writes the change into the copy. Git is not involved.
3. Other hooks record what the agent waits on you for, and at the end of a turn a small model keeps that list honest.
4. Tasks and records live in one SQLite database per machine. The agent uses them through the MCP server; you use the panel.

Details: Reference.

Privacy and cost

• Everything is stored on your machine: the copies in .claude/imprimatur/ in each repo (add it to .gitignore) and the database in your data folder.
• Model calls run through your own claude CLI and count against your Claude plan. Most use Haiku, and only where a rule is not enough: an edit's one-sentence description, closing answered asks, the end-of-turn audit, Scan history and placing an analysis.
• Once per repo with at least 10 records and no design documents, a background Sonnet run writes them (capped at $3). The best-practice review runs only when you ask, after a confirmation.
• Nothing else leaves your machine, except the issue tracker sync (GitHub with gh logged in, or Jira once you configure it), agnix fetched with npx for the audit, and the documentation the best-practice review fetches.
• IMPRIMATUR_DESCRIBE=off turns off edit descriptions. In an untrusted folder no model call, gh or repo script runs.

Links

• Reference: every feature, setup option, hook, MCP tool and setting.
• Design documents: Architecture, Product, Glossary.
• Changelog
• Contributing
• Security: report vulnerabilities privately.

License

MIT © 2026 Halil Ural