> ## Documentation Index
> Fetch the complete documentation index at: https://docs.myrmagent.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent Plugins Memory Bundle

> Give any Agent Plugins client (VS Code, Copilot, Kiro, Cursor, and more) persistent long-term memory backed by Myrm.

# Agent Plugins Memory Bundle

Myrm ships a portable [Agent Plugins](https://agent-plugins.org) bundle that attaches its long-term memory to any client supporting the standard — VS Code, GitHub Copilot, Kiro, Cursor, ChatGPT, Codex, and others. The bundle is transport-only: it points an MCP client at Myrm's memory server and teaches the consuming agent when and how to use it, with no per-client config files to edit.

## What you get

| File                          | Purpose                                                                                         |
| ----------------------------- | ----------------------------------------------------------------------------------------------- |
| `plugin.json`                 | Bundle manifest (name, version, description).                                                   |
| `mcp.json`                    | Streamable HTTP MCP server config pointing at Myrm's `/mcp` endpoint.                           |
| `skills/myrm-memory/SKILL.md` | Operating guidance for the consuming agent: when to recall, store, audit, and correct memories. |

## Generate a bundle

1. Open **Settings → Memory → Connect** (the Connect wizard).
2. Pick the Myrm agent whose memory you want to expose.
3. In the **Agent Plugins** card, choose whether to embed the token in the bundle or read it from the `MYRM_MCP_TOKEN` environment variable.
4. Click **Generate Agent Plugins bundle**, then download the whole bundle as **Download All (.zip)**, or copy each file individually. Unzip the archive — it already contains the correct layout — and place it into your client's plugin directory, then enable the plugin.

Keep the bundle structure above — `plugin.json` and `mcp.json` in the plugin root, `SKILL.md` under `skills/myrm-memory/`. The zip keeps this layout automatically; if you copy files one by one, recreate it by hand. Clients only load the skill file from its `skills/<name>/` subdirectory; a `SKILL.md` sitting flat in the root is silently ignored.

Embedding the token works out of the box but writes the credential into `mcp.json`, so keep such bundles out of version control. The default environment-variable mode keeps the bundle credential-free and safe to commit: most clients substitute `${MYRM_MCP_TOKEN}`, while some (VS Code, Cursor) use `${env:MYRM_MCP_TOKEN}`. For clients that cannot interpolate, turn on token embedding in the wizard — the token stays revocable at any time.

## Managing access

* The generated token is bound to the Myrm agent you selected. Only that agent's memory is reachable through the bundle.
* **Regenerate** the bundle in the wizard to mint a fresh token — the previous token is revoked immediately, so re-configure any client that still uses the old bundle.
* **Revoke** the connection in the wizard to invalidate the token immediately.
* **Test** the connection with the wizard's health check before relying on it.

## Running the Doctor

The wizard's **Doctor** button goes one step further than a live token test: it reads the MCP client's **actual config file on your machine** and verifies the bundle is wired correctly on the client side too.

| Diagnosis             | Meaning                                                                       | Severity        |
| --------------------- | ----------------------------------------------------------------------------- | --------------- |
| `verified`            | Config file found, `myrm-memory` entry present, Bearer token hash matches     | Healthy         |
| `token_valid`         | Token exists and matches the stored hash (no file to verify)                  | Healthy         |
| `token_env`           | Entry uses `${MYRM_MCP_TOKEN}` — the effective value can't be checked locally | Needs attention |
| `config_file_missing` | The configured path doesn't exist (check `~/.claude.json` etc.)               | Failed          |
| `entry_missing`       | File exists but no `myrm-memory` entry under the expected key                 | Failed          |
| `token_missing`       | Entry present but has no `Authorization: Bearer <token>` header               | Failed          |
| `token_mismatch`      | Token in the file doesn't match the stored hash — re-copy the snippet         | Failed          |
| `file_unreadable`     | File exists but isn't valid JSON/TOML                                         | Failed          |

* Both JSON (`json_mcp`) and TOML (`toml_mcp`) client formats are supported; `~` in paths is expanded automatically.
* Tokens are stored only as SHA-256 hashes — never in plaintext.
* Diagnoses are localized in 6 languages (EN / 中文 / 繁體 / 日本語 / 한국어 / DE) with an actionable fix hint for every code.
* The doctor only reads — it never modifies your client's config file.

## Self-hosting

The bundle is transport-only, so it also works against your own Myrm instance — local Web UI, desktop app, or cloud sandbox. Just point the `url` in `mcp.json` at your instance's `/mcp` endpoint and generate a token from its Connect wizard.