Install the Stashr agent skill

The MCP server and CLI give an agent access to your library. The Stashr agent skill teaches it the workflow: which transport to use, how to search cheaply before reading anything in full, when to look at images, and how to change your library safely. It is a small instruction file the agent loads when a task involves your saves, so you do not have to explain Stashr in every prompt.

Install

Install the skill from the official repository with the skills CLI:

bunx skills add DeckardGer/stashr-skills

Or with npx:

npx skills add DeckardGer/stashr-skills

This works with Claude Code and any other agent that supports the open skills format. Re-run the same command to update to the latest version.

The skill is instructions, not access. The agent still needs one connection to your library:

• A Stashr MCP connection, which the skill prefers when available.
• Or the Stashr CLI, signed in with stashr login or a scoped API key.

Agent access is part of Pro and included in every free trial.

What it teaches

Once installed, the agent follows the same conventions on every Stashr task:

• Pick one transport per session. Use a connected MCP server if present, otherwise fall back to the CLI once, instead of probing both repeatedly.
• Discover before reading. Start with compact, bounded search results, then fetch full content only for the few saves that matter to the task.
• Use image search for visual asks. Rank individual images by their AI captions and preview only a shortlist, instead of describing posts from text alone.
• Know the filter vocabulary. Valid platforms, content types, and author formats, including that X is twitter, so the agent never guesses values.
• Change things carefully. Save, tag, and archive only when asked, reuse your existing tag names, and require explicit confirmation before deleting a collection.
• Recover from errors. Read validation messages and correct the call, re-authenticate on expired sessions, and recognize an outdated CLI install.

When to use it

If you only chat with your library through Claude or ChatGPT over MCP, the connection alone is usually enough. The skill earns its keep with coding agents that work in long sessions, where consistent habits keep responses fast and token usage low, and with terminal agents driving the CLI, where it encodes flags and patterns the agent would otherwise rediscover each time.

Where to go next