diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..0668efb
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,29 @@
+# local-agent
+
+Optimization guide for running multiple LLMs on a single Apple Silicon GPU using llama.cpp router mode.
+
+## Key Configs
+
+- **Router:** `~/.hermes/models-router/` with symlinks, managed via launchd
+- **Plist:** `~/Library/LaunchAgents/com.jimmyg.llama-server-router.plist`
+- **INI:** `~/.hermes/llama-models.ini` (per-model threads: 35B t=14, 27B t=10)
+- **Old plists** disabled at `*.plist.disabled`
+- **Hermes providers** both point to `http://localhost:8085/v1`
+
+## Speeds
+
+| Model | Router Mode |
+|-------|-------------|
+| 35B MoE | 132 t/s |
+| 27B dense | 48 t/s |
+
+Model switch delay: ~3-4s. Slot-saving doesn't work with router mode.
+
+## Key Learnings
+
+- Don't `kill -9` launchd services — use `launchctl unload/load`
+- Batch: `512/128` not `2048/512` for single-user speed
+- Threads: MoE → 14, dense → 10
+- KV cache: Q4_0 > Q8_0 for bandwidth
+- MTP speculation: `--spec-draft-n-max 3` (~2.5x speedup)
+- Full lessons in `docs/04-lessons-learned.md`