[codex] Add MCP access, Chinese-first output mode, and dedicated ingest model

.gitignore

@@ -49,3 +49,8 @@ pnpm-debug.log*
 # Brainstorm assets (not tracked in source control)
 .superpowers/
 .claude/
+
+# Local Windows dev helpers
+LLM Wiki Dev.lnk
+launch-llm-wiki-dev.cmd
+tools/

mcp-server/README.md

@@ -0,0 +1,77 @@
+# LLM Wiki MCP Server
+
+Local stdio MCP server for the LLM Wiki desktop app. It wraps the app's local
+HTTP API at `http://127.0.0.1:19827/api/v1` so agents can discover and call
+LLM Wiki as named MCP tools instead of raw HTTP endpoints.
+
+## Requirements
+
+- Start the LLM Wiki desktop app.
+- Open the target project in the app.
+- Keep the local API reachable at `127.0.0.1:19827`.
+- Configure an LLM provider in LLM Wiki Settings before using
+  `llmwiki_chat`, `llmwiki_ingest_clip`, or `llmwiki_ingest_file`.
+
+## Run
+
+From this repository:
+
+```powershell
+npm run mcp:llmwiki
+```
+
+Equivalent direct command:
+
+```powershell
+node D:\Dev\llm_wiki\mcp-server\llmwiki-mcp.js
+```
+
+Optional environment variables:
+
+- `LLMWIKI_API_BASE`: defaults to `http://127.0.0.1:19827/api/v1`
+- `LLMWIKI_TIMEOUT_MS`: defaults to `60000` for ordinary calls; chat and
+  ingest calls keep a 30 minute timeout to match the desktop bridge.
+
+## Codex Registration
+
+```powershell
+codex mcp add llmwiki -- node D:\Dev\llm_wiki\mcp-server\llmwiki-mcp.js
+```
+
+Newly registered MCP servers may require a new Codex session before the tools
+appear.
+
+## Claude Desktop Example
+
+Add a server entry like this to your Claude Desktop MCP config:
+
+```json
+{
+  "mcpServers": {
+    "llmwiki": {
+      "command": "node",
+      "args": ["D:\\Dev\\llm_wiki\\mcp-server\\llmwiki-mcp.js"]
+    }
+  }
+}
+```
+
+## Tools
+
+- `llmwiki_status`: checks whether the local LLM Wiki API is reachable and
+  reports current project, recent projects, and capabilities.
+- `llmwiki_search`: quick ranked search across an LLM Wiki project.
+- `llmwiki_retrieve`: retrieves citation-ready context, relevant pages,
+  references, search hits, and graph expansions.
+- `llmwiki_chat`: asks LLM Wiki to answer using the app's configured LLM and
+  project knowledge base.
+- `llmwiki_graph`: returns the project knowledge graph; text summarizes counts
+  and `structuredContent` contains the full graph.
+- `llmwiki_ingest_clip`: saves text as a raw source in the active project and
+  queues it for ingest.
+- `llmwiki_ingest_file`: copies a local file into the active project's
+  `raw/sources` directory and queues it for ingest.
+
+Every tool returns a short text summary plus `structuredContent` containing the
+raw LLM Wiki API response. API failures return `isError: true` with a concrete
+hint to start the app and open the target project.