Claude Steward — 5DD Plan
Minimize Claude token usage without compromising efficiency. Keep your AI interactions governed, auditable, and cost-controlled.
Vision
Claude Steward is built around two pillars:
1. Token efficiency — squeeze the most out of every API call without degrading response quality. Smart routing, caching, and compression work together so you never pay for tokens you don't need.
2. Interaction governance — know what is being sent to Claude, when, and why. Lay the foundation for data hygiene, prompt auditing, and policy enforcement — so Claude fits inside your organization's boundaries, not around them.
Target: $5 a day (5DD plan) for a typical developer workload.
How it works
VS Code / Claude Code → localhost:8787 (Claude Steward proxy) → Anthropic API
| Technique |
What it does |
Typical saving |
| Model router |
Haiku classifies each request; simple ones stay on Haiku, complex ones escalate to Sonnet or Opus |
up to 10× cheaper per query |
| Semantic cache |
Returns cached response for repeated or similar prompts |
100% on cache hits |
| Prompt compression |
Strips whitespace, filler phrases, duplicate content |
5–15% |
| Context trimming |
Drops oldest messages when context exceeds budget |
20–50% on long conversations |
| Governance hooks (roadmap) |
Flag sensitive data, enforce prompt policies, log interactions |
— |
Setup (5 steps)
1. Install Node / npm
If you don't have npm, ask your support team to install Node.js.
2. Install dependencies
npm install
3. Build
npm run compile
4. Set the environment variable
Linux / macOS:
export ANTHROPIC_BASE_URL=http://localhost:8787
Windows (office / restricted environments):
Run set_env.bat — double-click or run from a terminal. Sets ANTHROPIC_BASE_URL permanently for your user session via setx.
5. Launch
Open the folder in VS Code and press F5.
Configuration
Open Settings (Ctrl+,) → search Claude Optimizer.
| Setting |
Description |
claudeOptimizer.proxyPort |
Proxy port (default: 8787) |
claudeOptimizer.enableCache |
Semantic response cache |
claudeOptimizer.cacheThreshold |
Similarity threshold 0–1 (default: 0.92) |
claudeOptimizer.enableModelRouter |
Auto-route to cheaper models |
claudeOptimizer.enablePromptCompression |
Compress prompts before sending |
Sensitive values via env vars:
ANTHROPIC_API_KEY, JIRA_API_TOKEN, GITHUB_TOKEN, AZURE_TENANT_ID, DATABRICKS_TOKEN, DATABASE_URL
Commands
| Command |
Description |
Claude Optimizer: Show Dashboard |
Live token savings dashboard |
Claude Optimizer: Toggle ON/OFF |
Pause/resume the proxy |
Claude Optimizer: Clear Cache |
Wipe the semantic cache |
Claude Optimizer: Optimize Selected Prompt |
Analyze selected text |
Debugging
Output panel → select "Claude Optimizer".
Key lines:
Proxy started on port 8787 — runningIncoming: model=... — request receivedStreaming: <model> rejected ... escalating to <model> — model escalation in progressCache HIT — saved a full round-trip
Roadmap
- [ ] Governance dashboard — audit log of all Claude interactions
- [ ] Sensitive data detection — flag PII/secrets before they leave your machine
- [ ] Prompt policy enforcement — block or warn on policy violations
- [ ] Cost alerts — notify when daily spend exceeds threshold
- [ ] Team mode — shared cache and governance rules across a team
The challenge
VS Code has no official API for intercepting Claude Code's calls. Claude Steward works by acting as a local proxy via ANTHROPIC_BASE_URL — a controlled man-in-the-middle on localhost. No extra API subscription required.
Contributing
PRs and issues welcome.
- One thing per PR
- No speculative abstractions
- If it's broken, open an issue before fixing
