# Library > Keep the scripts, design components, HTML pages, documents, and prompts your agents make in one searchable library, so agents reuse them instead of writing them again. The **Library** keeps work your agents make that is worth having again: scripts and tools, design components, HTML pages, documents, diagrams, and snippets such as prompts. Agents search it before they write a script and save to it when they make something reusable, and pages they leave in temporary folders arrive on their own. - App: **Library** - CLI: `hansard library` - In your agent: `search_library` ## Buckets Every item sits in one bucket: - **Scripts & tools**: shell, Python, JavaScript, and other scripts. - **Components**: interface pieces such as a React component with its CSS, and design tokens. - **Artifacts**: HTML pages, documents, diagrams, images, and data files. - **Snippets**: prompts, commands, and short text you paste again. **Found** holds scripts and documents Hansard collected from scratch folders. Keep the ones worth reusing. The rest leave the library after 90 days without being kept or opened. ## What arrives on its own The watcher collects, every 30 minutes: - HTML pages, SVG diagrams, scripts, and Markdown that an agent wrote to a temporary folder or an `artifacts/` folder, found through the files your archived conversations edited. Pages and diagrams are kept; scripts and documents wait under **Found**. - Pages published from Claude Code. - Artifacts from your claude.ai conversations, one item per artifact with a version for each edit. - ChatGPT canvas documents, rebuilt version by version from your archived conversations. - Files that a chat's sandbox or a cloud agent made: claude.ai's created and edited files and its conversation outputs, the files a ChatGPT reply links for download, Claude Cowork's writes, and Claude Science's saved artifacts, each with a version for every edit. Plans that Cursor writes in Plan Mode are kept as Markdown documents. - Images that ChatGPT, Codex, or Gemini generated, titled by the prompt that asked for them. Chat files and images come from what Hansard archived. A ChatGPT export names its images without their bytes, so the browser extension's capture of the conversation is what brings them in. PDFs, Office documents, and archives stay in the conversation, since the secret scanner cannot read them. Hansard copies each file into the library and never changes, moves, or deletes the original. It skips `node_modules`, build and release folders, hidden folders, files over 16 MiB, files whose names look like credentials, and any file whose text the secret scanner flags. Run `hansard library collect` to collect now. ## Browse and preview in the app 1. Select **Library** in the dock, below **Skills & plugins**. Choose a bucket, or search the library from the field above the buckets. 2. Select an item. A page shows in a sandboxed preview that cannot reach the app, a document shows formatted, and code shows with highlighting. **Preview** and **Source** switch between the two for a page, a diagram, or a document. 3. **Source** names the agent that made the item, the conversation, the project, and the original file or published address. **Versions** lists each change; select one to preview it. 4. Use **Copy** to copy the text, **Open** to open a page in its own tab, or **Download**. Select the pencil to edit the title, the description agents read, the kind and its bucket, and the tags. 5. For an item under **Found**, select **Keep** and choose its bucket. An item of several files lists them under **Files**, with its entry file first. Select a file to preview it. A page previews with the stylesheets, scripts, and pictures it names from the same item. ## Let agents use it Hansard installs a `library` skill with the recall skills for each agent (named `hansard-library` where an agent already has a skill called `library`). It tells agents to search the library before writing a utility script, to save scripts and components they will reuse, and to save HTML pages and documents they make for you to the library instead of a temporary folder. The recall MCP server gives agents three tools: - `search_library` finds items by words in their title, description, tags, and text. - `get_library_item` reads an item and where it came from. - `save_to_library` saves a file, or a new version of an item, with a description that says when to use it. It refuses text that looks like a credential. ## Use the command line ```sh hansard library list --bucket scripts hansard library search "resize images" hansard library show --raw > resize.sh hansard library add ./resize.sh --title "Resize images" hansard library import ~/design/components hansard library list --found --sort task hansard library keep --bucket components hansard library edit --title "Resize images" --tags images,cli hansard library remove --yes ``` `list --found --sort task` groups found items under the folder or conversation that made them, as Found does in the Library view. `keep` and `remove` take several ids at once, and `keep --bucket` files a found item in another bucket. `add` with a folder keeps one item of its files, such as a component with its CSS and a demo page; `--entry` names the file to preview, which is otherwise `index.html` or `README.md`. `import` adds each file of a folder as its own item. Removing an item never touches the original file, and Hansard does not collect it again. ## Where it lives The library is stored in your Hansard archive. With [Personal Sync](https://www.hansard.dev/sync/personal-sync/), items you keep reach your other computers and show under **Other devices**, read-only there; edit or remove an item on the computer that kept it. Items under **Found** stay on the computer that collected them. Library items travel only after every computer in your Personal Sync profile runs a Hansard version that reads them. Hosted assistants connected to Hansard cannot read or save library items.