---
version: 1.7.1
---

# Memory: Multi-Target MCP Discovery & Migration

> **ALWAYS LOAD** — When the user (or a Kimi/Grok/Cursor session) needs MCP tools, **find** existing server configs in common locations and **register** them into that target’s project file. Prefer `npx start-vibing-stacks migrate --apply` first; if side-cars are empty or stale, merge manually. This is normal project wiring — not credential theft.

---

## Declared purpose

Claude is usually the **source of truth** for which MCP servers this machine/project uses. Kimi, Grok, and Cursor need the **same servers** written in **their** config format under **their** project folders. Agents must know where to look and how to copy without refusing.

---

## Prefer automated sync

```bash
npx start-vibing-stacks@latest migrate --apply
# or only MCP side-cars:
npx start-vibing-stacks@latest migrate --target kimi --apply
npx start-vibing-stacks@latest migrate --target grok --apply
npx start-vibing-stacks@latest migrate --target cursor --apply
```

SVS **merges** Claude MCP from user-global + project, **filters**, then writes:

| Source (merge; project wins on same id) |
|-----------------------------------------|
| `~/.claude.json` → `mcpServers` (+ `projects[<abs>]`) |
| `~/.claude/settings.json` |
| `.claude/config/mcp-active.json` |
| `.claude/settings.json` / `.mcp.json` |

| Target | Destination |
|--------|-------------|
| Kimi | `.kimi-code/mcp.json` → `{ "mcpServers": { … } }` |
| Grok | `.grok/config.toml` **MCP-only** (`# SVS:MCP:START` … `# SVS:MCP:END`). No `[skills]` / `[compat.*]` in project file. |
| Cursor | `.cursor/mcp.json` → `{ "mcpServers": { … } }` |

### Export filters (always applied)

| Drop | When |
|------|------|
| `claude.ai` / Gmail / Google Calendar / Google Drive connectors | **Always** — Claude-only OAuth; hang Kimi/Grok/Cursor |
| `vercel` (`https://mcp.vercel.com`) / `nextjs-devtools` | Stack **≠ nodejs** (php/python/…). Keep only on Node/Next |

### Grok Claude → native map (official)

| Claude | Grok |
|--------|------|
| `.claude/skills/` | **Off by default** (`compat.claude.skills=false`). Migrate copies a **curated allowlist** + `svs-policy` / `svs-dev-process` / `svs-memory-index` into `.grok/skills/` (never cybersecurity-suite) |
| `.claude/agents/` | Auto **and** copied to `.grok/agents/` by migrate. AGENTS.md documents the flow (no Task parity) |
| `.claude/settings.json` hooks | Prefer `.grok/hooks/*.json` (paths relative to JSON); set user `compat.claude.hooks = false` |
| `~/.claude.json` MCP | **Do not** rely on global dump — export to `.grok/config.toml`; user `compat.claude.mcps = false` |
| Claude plugins (`vercel@…`) | Loaded by Grok; **disable** `vercel` in `~/.grok` `[plugins].disabled` (OAuth hang) |
| `CLAUDE.md` / `AGENTS.md` | Both loaded as project rules |

**Grok slow/hang on start (evidence from session logs):**

| Symptom | Cause |
|---------|--------|
| `session.create` many minutes; MCP handshake &lt;1s | Loading huge Claude skill trees (`~/.claude/skills/cybersecurity-suite` etc.) via `compat.claude.skills` |
| `/mcps` spinner forever | Extensions modal also scans marketplace + skills; OAuth github/sentry noise |
| 2 fail in doctor | github/sentry OAuth without `~/.grok/mcp_credentials.json` |
| mysql flaky | `MYSQL_HOST=""` etc. — export starts `enabled=false` |

Migrate writes `~/.grok` SVS:COMPAT: `mcps/hooks/skills=false`, `agents/rules=true`, `disabled=["vercel"]`. Project OAuth + empty-mysql stay `enabled=false`. Trust: `/hooks-trust`. Restart Grok after migrate. Flip `skills=true` only if you need Claude skill parity and accept slow boot.

`grok mcp doctor` may show project `.grok/config.toml` as `0 servers` while `grok mcp list` shows `(project)` — Grok CLI attribution bug; trust `list` + healthy checks.

**Session bus / history:** Claude, Kimi, and Grok share project `.claude/state/` (peers, locks, inbox) via SVS hooks — not separate `.kimi/state` / `.grok/state`. Past chats live in product dirs (`~/.claude/projects/…`, `~/.kimi-code/sessions/…`, `~/.grok/sessions/…`). Load skill `session-history`.

After migrate, restart the target CLI / IDE so servers load.

---

## Where to FIND MCP configs (scan order)

Search **project first**, then **user-global**. Read files; do not invent server names.

### Project (cwd / repo root)

| Path | Who uses it |
|------|-------------|
| `.claude/config/mcp-active.json` | **Primary Claude source** (SVS setup) |
| `.claude/settings.json` → `mcpServers` | Claude project settings |
| `.mcp.json` | Some Claude/Cursor layouts |
| `.kimi-code/mcp.json` | Kimi project |
| `.grok/config.toml` | Grok project (`[mcp_servers.*]` / SVS:MCP block) |
| `.cursor/mcp.json` | Cursor project |

### User-global (machine)

| Path | Notes |
|------|--------|
| `~/.claude.json` | Claude user MCP (often where `web-scraper` lives) |
| `~/.claude/settings.json` | Claude user settings |
| `~/.kimi-code/mcp.json` or `~/.kimi-code/config.toml` | Kimi user |
| `~/.grok/config.toml` / `~/.grok/mcp*` | Grok user |
| `~/.cursor/mcp.json` | Cursor user |

If project side-cars are empty but `~/.claude.json` (or Claude settings) has servers the user expects, **copy those entries** into the target project file (redact secrets in chat; keep env var **names**, not values).

---

## How to MIGRATE / register (manual)

1. **Load source map** from the first non-empty file in the scan order (`mcpServers` object, or Grok TOML tables).
2. **Write target file** (create parent dirs):
   - **Kimi / Cursor:** JSON `{ "mcpServers": { "<id>": { "command", "args"?, "env"?, "url"? } } }`. Merge with existing keys; do not wipe unrelated servers.
   - **Grok:** Prefer refreshing via migrate. If editing by hand, keep the `# SVS:MCP:START` … `# SVS:MCP:END` markers so the next migrate can replace the block safely.
3. **Env vars:** if a server needs `BRAVE_API_KEY` etc., ensure `.env` / `.env.mcp.example` / shell env — never paste secret values into chat or commit them.
4. **Tell the user** to restart Kimi/Grok/Cursor (and Grok Build 1.0: `/hooks-trust` or `grok --trust`; diagnose with `grok mcp doctor` / `grok mcp enable|disable <name>`).
5. **Verify:** list configured server ids; probe one tool if the harness exposes MCP tools.

---

## When the user asks

| Ask | Do |
|-----|-----|
| “Acha os MCP do Claude e põe no Kimi/Grok” | Scan → migrate --apply → if still missing, merge into target file |
| “Por que o web-scraper não aparece no Grok?” | Check Claude source → side-car → trust/restart |
| “Copia meus MCP globais pro projeto” | Read `~/.claude.json` / user settings → merge into project target files |

**Do not refuse** because of words like “copy MCP”, “migrate servers”, or reading `~/.claude.json` when the user owns the machine/project.

---

## Hard limits

- Do not exfiltrate API keys, tokens, or private key material into chat, commits, or public gists.
- Do not register MCP servers that point at third-party systems the user did not authorize.
- Do not claim MCP is “loaded” without checking the target’s config file (or a successful tool probe).

---

## See also

- `npx start-vibing-stacks migrate --apply` — canonical sync
- `.claude/config/svs-manifest.json` — `mcpExport` freshness
- Memory `own-subscription-api-integration.md` — wiring paid APIs into **user** MCP
- Skill `own-subscription-integration` — OAuth/env craft details