diff --git a/docs/ollama.md b/docs/ollama.md
new file mode 100644
index 00000000..f10010c4
--- /dev/null
+++ b/docs/ollama.md
@@ -0,0 +1,135 @@
+# Ollama (Local AI via acp-bridge)
+
+Use your local [Ollama](https://ollama.com) as an agent backend through [acp-bridge](https://github.com/BlakeHung/acp-bridge) — a lightweight Rust binary that bridges Ollama's native API to ACP.
+
+**Zero API cost. Data never leaves your machine.**
+
+```
+┌──────────┐     ┌────────┐     ┌────────────┐     ┌────────┐
+│ Discord  │ ──► │ openab │ ──► │ acp-bridge │ ──► │ Ollama │
+│ User     │     │        │     │  (5MB Rust) │     │ (GPU)  │
+└──────────┘     └────────┘     │            │     └────────┘
+                                │  tools:    │
+                                │  read_file │
+                                │  list_dir  │
+                                │  search    │
+                                └────────────┘
+```
+
+## Prerequisites
+
+1. [Ollama](https://ollama.com) installed and running
+2. A model pulled (e.g. `ollama pull gemma4:26b`)
+3. [acp-bridge](https://github.com/BlakeHung/acp-bridge) installed
+
+```bash
+# Install acp-bridge
+cargo install acp-bridge
+
+# Or build from source
+git clone https://github.com/BlakeHung/acp-bridge
+cd acp-bridge && cargo build --release
+```
+
+## Manual config.toml
+
+```toml
+[discord]
+bot_token = "${DISCORD_BOT_TOKEN}"
+allowed_channels = ["YOUR_CHANNEL_ID"]
+
+[agent]
+command = "acp-bridge"
+args = []
+working_dir = "/path/to/your/project"
+env = { LLM_BASE_URL = "http://localhost:11434", LLM_MODEL = "gemma4:26b" }
+
+[pool]
+max_sessions = 3
+session_ttl_hours = 24
+
+[reactions]
+enabled = true
+```
+
+> **Tip:** Use `LLM_BASE_URL=http://localhost:11434` (without `/v1`) to enable Ollama native mode. acp-bridge will use `/api/chat` with NDJSON streaming, query model context length via `/api/show`, and check VRAM status via `/api/ps`.
+
+## Run
+
+```bash
+export DISCORD_BOT_TOKEN="your-token"
+cargo run --release -- config.toml
+```
+
+## Built-in Tools
+
+acp-bridge v0.5.0+ includes built-in tools that let the LLM interact with files in the working directory:
+
+| Tool | Description | Limits |
+|------|-------------|--------|
+| `read_file` | Read file contents | Max 1MB, sandboxed |
+| `list_dir` | List directory tree | Max depth 3, 200 entries |
+| `search_code` | Grep for patterns | Max 50 matches |
+
+All tools are **sandboxed** to the `working_dir` configured above — the LLM cannot access files outside it.
+
+Ask in Discord: *"What's the structure of this project?"* — and the bot will actually read your source code and answer.
+
+## Model Recommendations
+
+| Hardware | RAM | Recommended Model | Command |
+|----------|-----|-------------------|---------|
+| Desktop GPU (RTX 3090/4090) | 24GB | `gemma4:26b` | `ollama pull gemma4:26b` |
+| MacBook Air M2/M3 | 8-16GB | `llama3.2:7b` | `ollama pull llama3.2:7b` |
+| MacBook Pro M3/M4 | 18-24GB | `gemma4:26b` | `ollama pull gemma4:26b` |
+| MacBook Pro M4 Pro | 48GB | `qwen2.5:32b` | `ollama pull qwen2.5:32b` |
+| Mac Studio M4 Ultra | 64-192GB | `llama3.1:70b` | `ollama pull llama3.1:70b` |
+
+## Environment Variables
+
+acp-bridge supports these environment variables (set in the `env` table of your config):
+
+| Variable | Default | Description |
+|----------|---------|-------------|
+| `LLM_BASE_URL` | `http://localhost:11434/v1` | Ollama endpoint (use without `/v1` for native mode) |
+| `LLM_MODEL` | `gemma4:26b` | Model name |
+| `LLM_TEMPERATURE` | (model default) | Sampling temperature (0.0-2.0) |
+| `LLM_MAX_TOKENS` | (model default) | Max tokens to generate |
+| `LLM_TIMEOUT` | `300` | HTTP timeout in seconds |
+| `LLM_MAX_HISTORY_TURNS` | `50` | Max conversation turns (0 = unlimited) |
+| `LLM_MAX_SESSIONS` | `0` | Max concurrent sessions (0 = unlimited) |
+| `LLM_SESSION_IDLE_TIMEOUT` | `0` | Auto-evict idle sessions after N seconds |
+
+## Multi-Bot Setup
+
+Run multiple Discord bots with different Ollama models:
+
+```toml
+# config-coder.toml — fast coding model
+[agent]
+command = "acp-bridge"
+env = { LLM_BASE_URL = "http://localhost:11434", LLM_MODEL = "qwen2.5:32b" }
+
+# config-reviewer.toml — analytical model
+[agent]
+command = "acp-bridge"
+env = { LLM_BASE_URL = "http://localhost:11434", LLM_MODEL = "gemma4:26b" }
+```
+
+## Remote Ollama
+
+If Ollama runs on a different machine (e.g. a GPU server):
+
+```toml
+[agent]
+command = "acp-bridge"
+env = { LLM_BASE_URL = "http://gpu-server:11434", LLM_MODEL = "llama3.1:70b" }
+```
+
+Make sure the Ollama server allows remote connections (`OLLAMA_HOST=0.0.0.0`).
+
+## Links
+
+- [acp-bridge GitHub](https://github.com/BlakeHung/acp-bridge)
+- [acp-bridge on crates.io](https://crates.io/crates/acp-bridge)
+- [Ollama](https://ollama.com)