diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..bc5f0ba
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,94 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Commands
+
+```bash
+# Build
+go build ./...
+
+# Run main server
+go run cmd/server-bot/main.go
+
+# Test all packages
+go test ./...
+
+# Test a single package
+go test ./internal/history/...
+
+# Run a specific test
+go test -run TestName ./internal/server/...
+
+# Docker (recommended for full stack)
+docker-compose up
+```
+
+## Architecture
+
+hairy-botter is a **modular AI chatbot server** built with Firebase Genkit. HTTP clients (Telegram, CLI, WhatsApp, etc.) communicate with the main server over a REST API. The server handles sessions, RAG retrieval, history, and delegates to Genkit for LLM generation with agentic tool use.
+
+### Request flow
+
+```
+Client (HTTP POST /message)
+  → internal/server   — extracts session ID, files, message
+  → internal/ai/agent — reads history, retrieves RAG context, calls genkit.Generate()
+                         Genkit handles the agentic loop (tool calls, MCP, Google Search)
+  → internal/history  — saves updated conversation (resp.History())
+  → returns text response
+```
+
+### Key packages
+
+| Package | Role |
+|---|---|
+| `cmd/server-bot` | Entry point — wires all dependencies, reads env vars |
+| `internal/ai/agent` | Core AI logic: history + RAG + Genkit generate |
+| `internal/ai/gemini` | Genkit plugin setup (model, embedder, Google Search config) |
+| `internal/ai/gemini-embedding` | Adapter: Genkit embedder → chromem `EmbeddingFunc` |
+| `internal/server` | Chi HTTP router, session cookies, multipart file handling |
+| `internal/history` | JSON session history; summarizes when length exceeds threshold |
+| `internal/rag` | chromem-go vector DB — loads `bot-context/` docs on startup |
+| `cmd/server-mcp` | Standalone MCP protocol server |
+| `cmd/server-mcp-skills` | Sandboxed skills MCP server (file I/O, shell commands) |
+
+### MCP tool pattern
+
+Tools are registered with:
+```go
+genkit.DefineTool(g, name, desc, fn, ai.WithInputSchema(schema))
+```
+The tool `fn` signature is `func(tc *ai.ToolContext, input any) (any, error)` — `tc` embeds `context.Context`. Genkit handles the agentic tool-call loop automatically. **Do not call `gemini.New()` twice with overlapping tool names** — tool registration is global per Genkit instance.
+
+### Session & history
+
+- Session ID comes from a cookie or `X-User-ID` header.
+- History is stored as `[]*ai.Message` (genkit type) in `history-gemini/<sessionID>.json`.
+- Field name in JSON is `content` (not `parts` — old genai format is incompatible).
+- When history length hits `HISTORY_SUMMARY` (default 20), a summary is generated and replaces the messages.
+
+### Configuration (environment variables)
+
+| Variable | Default | Purpose |
+|---|---|---|
+| `GEMINI_API_KEY` | — | **Required.** Google AI API key |
+| `ADDR` | `:8080` | Listen address |
+| `GEMINI_MODEL` | `gemini-2.5-flash` | Model name |
+| `MCP_SERVERS` | — | Comma-separated MCP HTTP server URLs |
+| `GEMINI_SEARCH_DISABLED` | `false` | Disable Google Search grounding |
+| `HISTORY_SUMMARY` | `20` | Summarization threshold (message count) |
+| `LOG_LEVEL` | `info` | `debug`/`info`/`warn`/`error` |
+| `CORS_ALLOWED_ORIGIN` | `*` | CORS origin |
+
+**File-based config** (loaded from working directory at startup):
+- `personality.txt` — system prompt (supports both `{"text":"..."}` and `{"parts":[...]}` JSON formats)
+- `bot-context/` — documents auto-indexed into RAG vector DB
+
+### Genkit specifics
+
+- `genkit.Init` panics on failure (does not return error).
+- For known Gemini models: `ga.DefineModel(g, name, nil)`.
+- For unknown models: provide `*ai.ModelOptions{Supports: &googlegenai.Multimodal}`.
+- Google Search grounding is Gemini-specific and passed via `ai.WithConfig(&genai.GenerateContentConfig{Tools: ...})`.
+- The genkit fork used is `gerifield/genkit/go v1.5.0-fix` (replace directive in go.mod).