diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 000000000..96a64d920
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,114 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Development Commands
+
+### Full app
+
+```bash
+# Run backend + frontend with Docker Compose
+docker-compose up
+```
+
+Docker exposes the FastAPI backend on `${PORT:-8001}` and the Next.js frontend on `3000`. Runtime data, repository clones, embeddings, and wiki cache are persisted under `~/.adalflow`.
+
+### Backend API
+
+```bash
+# Install backend dependencies from the repository root
+python -m pip install poetry==2.0.1 && poetry install -C api
+
+# Start the FastAPI server on PORT or 8001
+python -m api.main
+
+# Alternate helper used by run.sh
+uv run -m api.main
+```
+
+The backend loads `.env` from the repository root. `GOOGLE_API_KEY` and `OPENAI_API_KEY` are the common baseline keys; other providers are enabled by their matching environment variables and `api/config/*.json` entries. `PORT` defaults to `8001`.
+
+### Frontend
+
+```bash
+# Install JavaScript dependencies
+npm install
+
+# Start Next.js on port 3000
+npm run dev
+
+# Production build and start
+npm run build
+npm run start
+
+# Lint
+npm run lint
+```
+
+The frontend uses `SERVER_BASE_URL` to reach the backend, defaulting to `http://localhost:8001`.
+
+### Tests
+
+```bash
+# All tests via the project runner
+python tests/run_tests.py
+
+# Test categories
+python tests/run_tests.py --unit
+python tests/run_tests.py --integration
+python tests/run_tests.py --api
+
+# Single test file examples
+python tests/unit/test_google_embedder.py
+python tests/api/test_api.py
+
+# Pytest directly, when useful
+pytest tests/unit/test_google_embedder.py
+pytest tests/unit/test_google_embedder.py::test_name
+```
+
+Some tests require API keys (`GOOGLE_API_KEY`, `OPENAI_API_KEY`) or a running backend for API tests.
+
+## Architecture Overview
+
+DeepWiki is split into a Next.js frontend and a FastAPI backend. The frontend collects repository details, provider/model selection, tokens, language, and include/exclude filters, then calls backend endpoints to clone/index repositories, generate wiki content, and stream Ask responses.
+
+### Frontend (`src/`)
+
+- `src/app/page.tsx` is the landing/configuration flow. It parses GitHub/GitLab/Bitbucket/local inputs, stores per-repository UI settings in `localStorage`, and routes to repository wiki pages.
+- `src/app/[owner]/[repo]/page.tsx` is the main wiki UI. It builds request bodies with repo credentials, provider/model selection, language, and file filters, then renders generated wiki pages and Ask interactions.
+- `src/app/api/*/route.ts` files are Next.js proxy/fallback routes. `next.config.ts` also rewrites selected frontend API paths to the backend using `SERVER_BASE_URL`.
+- `src/utils/websocketClient.ts` is the primary chat streaming client for Ask; `src/app/api/chat/stream/route.ts` keeps an HTTP streaming fallback.
+- `src/contexts/LanguageContext.tsx` and `src/messages/*.json` provide localized UI strings.
+
+### Backend (`api/`)
+
+- `api/main.py` loads environment variables, configures logging, patches development file watching, checks common API keys, and starts Uvicorn with `api.api:app`.
+- `api/api.py` defines FastAPI models and routes for auth, model config, wiki cache, export, repository processing, local repository structure, health, and streaming/chat endpoints.
+- `api/config.py` loads `api/config/generator.json`, `embedder.json`, `repo.json`, and `lang.json`, replaces `${ENV_VAR}` placeholders, and maps provider/client names to implementation classes.
+- `api/data_pipeline.py` clones repositories, reads source/doc files with include/exclude filters, splits content, creates embeddings, and stores indexes in AdalFlow local databases.
+- `api/rag.py` builds the retrieval-augmented Ask pipeline using AdalFlow retrievers, the configured embedder, provider-specific generator clients, and a small in-memory conversation history wrapper.
+- Provider clients (`openai_client.py`, `openrouter_client.py`, `bedrock_client.py`, `azureai_client.py`, `dashscope_client.py`, Google/Ollama integration) adapt each LLM provider to the generator/embedder config model.
+
+### Configuration and storage
+
+- `api/config/generator.json` defines generation providers, model lists, client classes, defaults, and model kwargs.
+- `api/config/embedder.json` defines the active embedding backend; `DEEPWIKI_EMBEDDER_TYPE` selects between OpenAI, Google, Ollama, and Bedrock-style embedder configs.
+- `api/config/repo.json` controls repository limits and file filters.
+- `DEEPWIKI_CONFIG_DIR` can point the backend at an alternate config directory without changing code.
+- Local runtime state is under `~/.adalflow`: cloned repositories, vector databases, and wiki cache. Docker Compose mounts this path into the container.
+
+### Data flow
+
+1. The user enters a repository URL/path and options in the Next.js UI.
+2. The backend clones or reads the repository, applying file filters from the request and `repo.json`.
+3. The data pipeline chunks files and creates embeddings via the configured embedder.
+4. Wiki generation and Ask responses use the selected provider/model through `api/config.py` and provider client classes.
+5. Generated wiki structures are cached locally and exposed back to the frontend for navigation, export, and subsequent Ask requests.
+
+## Notes for Future Agents
+
+- Keep frontend proxy behavior in sync between `next.config.ts`, `src/app/api/*/route.ts`, and backend route names in `api/api.py`.
+- When changing model/provider behavior, update `api/config/generator.json` or `api/config/embedder.json` first, then verify the matching client class and environment variables in `api/config.py`.
+- Be careful with tokens in repository clone paths and logs; `api/data_pipeline.py` sanitizes clone errors and logs the original repository URL instead of tokenized URLs.
+- Backend tests and integration flows may call external LLM/embedding services. Prefer focused unit tests or mocked clients when changing pure parsing/configuration behavior.
\ No newline at end of file