claude-omlx

gistfile1.txt

#!/usr/bin/env bash
# claude-omlx — run Claude Code as a UI for a local omlx server.
#
# Builds on the env-var profile from https://github.com/vitorallo/claude-code-local
# (the 15 problems it documents), but keeps omlx as the backend instead of
# vllm-mlx (omlx is forked from vllm-mlx, adds paged-SSD KV cache, and is
# actively maintained for Claude Code).
#
# Token lives in macOS Keychain — service "claude-omlx", account "omlx":
#   security add-generic-password -s claude-omlx -a omlx -w '<token>' -U
#
# Override at call time:
#   OMLX_BASE_URL=http://127.0.0.1:8001 OMLX_MODEL=other claude-omlx

set -euo pipefail

# Drop any real Anthropic creds inherited from the parent shell so they
# cannot leak to the local server.
unset ANTHROPIC_API_KEY ANTHROPIC_AUTH_TOKEN || true

if [[ -z "${OMLX_TOKEN:-}" ]]; then
  if ! OMLX_TOKEN="$(security find-generic-password -s claude-omlx -a omlx -w 2>/dev/null)"; then
    echo "claude-omlx: token not in Keychain (service=claude-omlx, account=omlx)." >&2
    echo "  add it with: security add-generic-password -s claude-omlx -a omlx -w '<token>' -U" >&2
    exit 1
  fi
fi

# Defaults to http://127.0.0.1:8000
OMLX_BASE_URL="${OMLX_BASE_URL:-http://127.0.0.1:8000}"
OMLX_MODEL="${OMLX_MODEL:-Qwen3.6-35B-A3B-6bit}"

# Anthropic SDK routing. Set AUTH_TOKEN only — also setting API_KEY triggers
# Claude Code's "Auth conflict" warning and can shadow the token at runtime.
export ANTHROPIC_BASE_URL="$OMLX_BASE_URL"
export ANTHROPIC_AUTH_TOKEN="$OMLX_TOKEN"

# Route every Claude Code model tier at the local model. Without these,
# Claude Code calls claude-opus-*, claude-sonnet-*, claude-haiku-* by
# canonical name — your local server 404s on those and the UI falls back
# to displaying "Sonnet".
export ANTHROPIC_MODEL="$OMLX_MODEL"
export ANTHROPIC_DEFAULT_OPUS_MODEL="$OMLX_MODEL"
export ANTHROPIC_DEFAULT_SONNET_MODEL="$OMLX_MODEL"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="$OMLX_MODEL"
export CLAUDE_CODE_SUBAGENT_MODEL="$OMLX_MODEL"

# Disable cloud-only behaviors that hang or break against a local server.
export CLAUDE_CODE_ATTRIBUTION_HEADER=0     # otherwise KV cache invalidates per request
export DISABLE_PROMPT_CACHING=1
export DISABLE_NON_ESSENTIAL_MODEL_CALLS=1
export DISABLE_AUTOUPDATER=1
export DISABLE_TELEMETRY=1
export DISABLE_ERROR_REPORTING=1

# Empty MCP config so plugins don't flood the local model with hundreds of
# tool defs (a 35B-A3B MoE with 3B active params collapses into bad tool
# calling under that load, regardless of quant).
MCP_CONFIG_DIR="${XDG_CONFIG_HOME:-$HOME/.config}/claude-omlx"
MCP_CONFIG_FILE="$MCP_CONFIG_DIR/mcp-local.json"
mkdir -p "$MCP_CONFIG_DIR"
[[ -f "$MCP_CONFIG_FILE" ]] || printf '{"mcpServers":{}}\n' > "$MCP_CONFIG_FILE"

exec claude --strict-mcp-config --mcp-config "$MCP_CONFIG_FILE" "$@"