Claude with Qwen3.6-A3B

run-local-model.sh

#!/usr/bin/env bash
# Run Claude Code against a local llama-server.
#
# Assumes you've already started llama-server yourself on $PORT (default 8080).
# llama.cpp's /v1/messages endpoint speaks the Anthropic Messages API directly,
# so Claude Code can talk to it with no proxy. (Added in llama.cpp PR #17570.)
#
# Usage:
#   ./run-claude-local.sh                 # connect to 127.0.0.1:8080
#   PORT=9090 ./run-claude-local.sh       # different port
#   HOST=192.168.1.5 ./run-claude-local.sh
#
# NOTE: CLAUDE_CODE_ATTRIBUTION_HEADER must be set in ~/.claude/settings.json
# (shell exports are ignored for that var when ANTHROPIC_BASE_URL is set):
#   { "env": { "CLAUDE_CODE_ATTRIBUTION_HEADER": "0" } }

set -euo pipefail

PORT="${PORT:-8080}"
HOST="${HOST:-127.0.0.1}"
ALIAS="${ALIAS:-local-model}"

command -v claude >/dev/null || { echo "claude not found in PATH" >&2; exit 1; }

if ! curl -fsS "http://$HOST:$PORT/v1/models" >/dev/null 2>&1 \
   && ! curl -fsS "http://$HOST:$PORT/health" >/dev/null 2>&1; then
  echo "No server responding at http://$HOST:$PORT — start llama-server first." >&2
  exit 1
fi

export ANTHROPIC_BASE_URL="http://$HOST:$PORT"
export ANTHROPIC_AUTH_TOKEN="${ANTHROPIC_AUTH_TOKEN:-sk-local-dummy}"
export ANTHROPIC_MODEL="${ANTHROPIC_MODEL:-$ALIAS}"

# Route the small/fast model to the same local server so topic naming,
# summarization, and other Haiku-tier calls don't leak to api.anthropic.com.
export ANTHROPIC_SMALL_FAST_MODEL="${ANTHROPIC_SMALL_FAST_MODEL:-$ALIAS}"

# Kill telemetry/statsig/update probes that hit api.anthropic.com regardless
# of ANTHROPIC_BASE_URL and inject volatile state into requests.
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1
export DISABLE_TELEMETRY=1
export DISABLE_ERROR_REPORTING=1
export DISABLE_AUTOUPDATER=1
export DISABLE_BUG_COMMAND=1
export CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1

echo "Launching Claude Code → $ANTHROPIC_BASE_URL (model: $ANTHROPIC_MODEL)"
exec claude "$@"

settings.json

{
  "env": {
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  },
  "model": "opus",
  "statusLine": {
    "type": "command",
    "command": "input=$(cat); printf \"⏰ %s | 👤 %s | 🤖 %s\" \"$(date +%H:%M:%S)\" \"$(whoami)\" \"$(echo \"$input\" | jq -r '.model.display_name')\""
  },
  "enabledPlugins": {
    "swift-lsp@claude-plugins-official": true,
    "clangd-lsp@claude-plugins-official": true,
    "rust-analyzer-lsp@claude-plugins-official": true
  },
  "effortLevel": "medium"
}