Memclaw v0.2.0 — 81% cheaper with raw Anthropic API
Replaced the Claude Agent SDK with a direct Anthropic API integration. The SDK silently injected ~20k tokens of unused built-in tool definitions on every call — removing it drops
real-world cost from $0.13 to $0.025 per query.
Benchmark results (claude-sonnet-4-6)
| Metric | v0.1.1 (SDK) | v0.2.0 (Raw API) |
|---|---|---|
| Avg cost/query | $0.132 * | $0.025 |
| Avg latency | 11.2s | 6.2s |
| Input tokens/call | ~20,000+ | ~7,400 |
| Cost reduction | — | ~81% |
- SDK costs measured from 16 production Telegram bot queries. Raw API costs from a 10-query benchmark.
What changed
- claude-agent-sdk dependency removed — no more Claude Code subprocess spawned per session
- Hand-rolled agentic loop — simple while-loop with tool dispatch, only our 8 tools are sent to the API
- Token-efficient tools beta enabled via anthropic-beta header
- Per-query cost logging — every call now logs exact input/output token counts and dollar cost
- Tools extracted to tools.py — agent.py cut from 689 to 310 lines
No breaking changes
Same CLI commands, same Telegram bot, same ~/.memclaw/ file layout, same env vars.
Upgrading
pip install --upgrade memclaw
pip uninstall claude-agent-sdk # no longer needed