ypollak2/llm-router

What's new in v4.0.0

Tool Slim Mode (LLM_ROUTER_SLIM)

Register only 12 or 4 tools at startup instead of all 43 — saves 5,000–8,000 context tokens per session. Gating happens at registration time, so unused tools never appear in Claude's tool list.

export LLM_ROUTER_SLIM=routing   # 12 tools — recommended
export LLM_ROUTER_SLIM=core      # 4 tools — maximum savings

Real-time Session Spend Meter (llm_session_spend)

Tracks API spend in a flat JSON file updated after every call — no Python needed in hook scripts. Anomaly detection fires when spend exceeds LLM_ROUTER_ANOMALY_THRESHOLD ($0.50 default) in under 10 minutes.

Correction Learning (llm_reroute)

Route corrections are stored in SQLite; llm_route explain mode automatically lowers confidence (−15% per correction), closing the feedback loop without manual tuning.

Quickstart Wizard (llm-router quickstart)

5-step interactive setup: detect hosts → configure API keys → install MCP config → test first call → show savings projection.

Cost-threshold Escalation (LLM_ROUTER_ESCALATE_ABOVE)

Block calls estimated above a per-call USD threshold before the model loop starts. Use llm_approve_route to approve/reject pending decisions.

Context Analyzer (llm_fs_analyze_context)

Reads up to 20 key project files, routes summarization to a cheap model, and persists a context summary for offline use.

Doctor Host Checks (llm-router doctor --host)

Check host-specific MCP config, hook installation, and connectivity:

llm-router doctor --host claude
llm-router doctor --host vscode
llm-router doctor --host all

Upgrade

pip install --upgrade claude-code-llm-router && llm-router install

VS Code extension (backlogged to v4.1)

The VS Code/Cursor status bar extension was assessed as highest-risk and moved to the v4.1 backlog.