Initial commit: llm-router — smart OpenAI/Anthropic/Ollama router

A stdlib pre-router in front of Ollama with LiteLLM backend:
- auto model selection by content/tools/modality, with fallbacks
- OpenAI /v1, Anthropic /v1/messages, and Ollama-native /api/* endpoints
- Whisper-shaped /v1/audio/transcriptions + in-chat audio
- key-based fleet policies (e.g. force a client onto uncensored models)
- optional Bearer auth; launchd/systemd service install
- benchmark harnesses (speed, quality, agentic tool use) with sample results

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

scripts/make-context-variant.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# Create a copy of an Ollama model with a fixed context window (num_ctx).
+# Useful because Ollama's default context is small and large prompts get truncated.
+#
+# Usage:  ./scripts/make-context-variant.sh <base-model> <num_ctx> [new-tag]
+# Example: ./scripts/make-context-variant.sh glm-4.7-flash 131072 glm-4.7-flash:128k
+#
+# Memory note: KV cache grows with num_ctx. On a 36 GB unified-memory Mac, keep
+# the model + KV under ~27 GB (the default GPU wired ceiling) to avoid swapping.
+set -euo pipefail
+BASE="${1:?usage: make-context-variant.sh <base-model> <num_ctx> [new-tag]}"
+NCTX="${2:?need a num_ctx value, e.g. 65536}"
+TAG="${3:-${BASE%%:*}:${NCTX}ctx}"
+
+TMP="$(mktemp -t Modelfile.XXXXXX)"
+printf 'FROM %s\nPARAMETER num_ctx %s\n' "$BASE" "$NCTX" > "$TMP"
+ollama create "$TAG" -f "$TMP"
+rm -f "$TMP"
+echo ">> created $TAG (num_ctx=$NCTX from $BASE)"
+ollama show "$TAG" | grep -i num_ctx || true

scripts/pull-models.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+# Pull a model fleet into Ollama. EDIT the MODELS list to taste, then run.
+# Continues past failures; prints a summary. Ollama resumes partial downloads.
+#
+# The names here must line up with the routing map in router.py (M / UNCENSORED).
+# If you change models, update router.py accordingly (see docs/ROUTING.md).
+set -u
+
+MODELS=(
+  # --- general / small ---
+  "qwen3:8b"
+  "qwen3:14b"
+  # --- reasoning / agentic ---
+  "qwen3.6:35b-a3b"
+  "gpt-oss:20b"
+  "glm-4.7-flash"
+  # --- coding ---
+  "qwen3-coder:30b"
+  # --- vision / OCR ---
+  "qwen3-vl:8b"
+  "qwen3-vl:30b-a3b-instruct"
+  # --- multimodal (vision + audio) ---
+  "gemma4:e4b"          # ROUTER_AUDIO_MODEL default — needed for /v1/audio/*
+  "gemma4:12b"
+  "gemma4:26b"
+
+  # --- example: a GGUF straight from Hugging Face (see docs/ROUTING.md) ---
+  # "hf.co/USER/REPO:Q4_K_M"
+)
+
+ok=(); fail=(); i=0; total=${#MODELS[@]}
+for m in "${MODELS[@]}"; do
+  i=$((i+1))
+  echo "=== [$i/$total] pulling $m ==="
+  if ollama pull "$m"; then ok+=("$m"); else echo "!! failed: $m"; fail+=("$m"); fi
+done
+
+echo ""; echo "done. ${#ok[@]}/$total succeeded."
+[ ${#fail[@]} -gt 0 ] && { echo "failed:"; printf '  - %s\n' "${fail[@]}"; }

scripts/setup.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+# One-time setup: create the Python 3.12 venv and install LiteLLM + Pillow.
+# The pre-router itself needs no venv (stdlib only); this is for the backend proxy.
+set -euo pipefail
+cd "$(dirname "$0")/.."
+
+PY="${PYTHON:-python3.12}"
+if ! command -v "$PY" >/dev/null 2>&1; then
+  # common macOS Homebrew fallback
+  [ -x /opt/homebrew/opt/python@3.12/bin/python3.12 ] && PY=/opt/homebrew/opt/python@3.12/bin/python3.12
+fi
+command -v "$PY" >/dev/null 2>&1 || { echo "Need Python 3.12. Set \$PYTHON, or:"; \
+  echo "  macOS:  brew install python@3.12"; echo "  Linux:  install python3.12"; exit 1; }
+
+echo ">> using $($PY --version) at $(command -v "$PY" 2>/dev/null || echo "$PY")"
+[ -d .venv ] || "$PY" -m venv .venv
+.venv/bin/pip install -q --upgrade pip
+.venv/bin/pip install -q -r requirements.txt
+.venv/bin/python -c "import litellm, PIL" && echo ">> deps OK (litellm, pillow)"
+echo ""
+echo "Next:"
+echo "  ./scripts/pull-models.sh   # pull a model fleet (edit the list first)"
+echo "  ./run.sh                   # start LiteLLM + router"