agent-wiki — Claude Code plugin

agent-wiki gives agents private markdown project memory over MCP. This public repository distributes version 0.3.0, including the explicitly invoked /agent-wiki:setup skill and a remote server at https://getagentwiki.com/mcp. It bundles no credentials, dependencies, or automatic write hooks.

Install

First inspect /mcp in Claude Code. If agent-wiki already works through a direct server or a claude.ai connector, preserve that connection and read Existing connections before installing.

Run these commands in your shell:

claude plugin marketplace add animal-ltd/agent-wiki-plugin
claude plugin install agent-wiki@agent-wiki --scope user

The interactive equivalents are /plugin marketplace add animal-ltd/agent-wiki-plugin and /plugin install agent-wiki@agent-wiki; choose the intended scope in the plugin manager.

• user: the skill and bundled server are available to you across projects.
• project: shared with collaborators through this repository's Claude settings.
• local: only you, in this project.

Use --scope project or --scope local only when intended. Installing in a project directory with --scope user still installs across projects. Installation scope does not grant wiki access. The bundled endpoint is account-wide; do not replace an existing single-wiki connection without choosing that broader connection yourself.

Restart Claude Code. In /plugin, check Installed for enabled agent-wiki version 0.3.0 and inspect Errors if /agent-wiki:setup is missing. A client that requests /reload-plugins supports that alternative; follow its reload instructions. Shell installation does not update an already-running conversation.

Authenticate, then run setup

Installation only loads the skill and server. Open /mcp, select the active agent-wiki server, and choose Authenticate or Re-authenticate when needed. Complete sign-in and approval in the browser and return to the waiting client. Never share credentials or OAuth callback URLs. For a direct server named agent-wiki, Claude Code v2.1.186 or later also supports claude mcp login agent-wiki. A non-interactive prompt cannot complete interactive OAuth.

Once /mcp shows the connection and tools, invoke:

/agent-wiki:setup

You may supply a wiki name or id. The skill confirms access with list_wikis, asks if selection is ambiguous, checks the wiki's state, and reads at most three orientation pages. It configures private project instructions and a read-only retrieval hook using that existing connection. Native hooks require Claude Code 2.1.282 or newer. Before enabling retrieval, it explains that each current prompt, including pasted text, is sent transiently for retrieval and lets you choose instructions-only instead. The hook adds no separate transcript, local-path, or credential fields, but all text included in the prompt is sent. Agent-wiki does not retain that prompt text in its logs; client conversation history is separate. Existing settings, permissions, and other hooks are preserved.

Configuration does not require saving a new decision. If one is needed, the assistant shows the exact note and destination and waits for your approval. It uses a stable write request ID, preserves current versions, verifies the saved revision, and searches the result. Existing decisions are not saved again. No examples are saved as decisions.

If credentials expire, return to /mcp and re-authenticate the selected connection. If permission is read-only, access is missing, a tool call is refused, or you cancel, setup reports what remains incomplete and does not bypass the failure. A rerun checks for an already-saved decision rather than duplicating it. A save followed by failed verification is reported as unverified, not complete.

Existing connections

/mcp identifies server origins and hidden duplicates. Direct definitions can take preced

... View full README on GitHub