I stared at my Anthropic bill. $200. For one month. Of Claude Code.
Not because I was doing anything special. Just... regular coding. But here's the thing nobody tells you:
Claude Code sends 27 tool definitions with every single request.
27 JSON schemas. Every turn. Even if you're just fixing a typo in a CSS file.
I was paying to send GitHub tools, Context tools, IDE tools, agent tools... when all I needed was Read
, Edit
, and Grep
.
So I built a proxy. A dumb, simple proxy that sits between my editor and the API. It reads my message, figures out what tools I actually need, and strips the rest before it reaches the model.
Cost went from $200 to $50. Same code. Same quality.
Here's how.
Every AI coding tool works the same way. They have a bunch of built-in tools. File reading, bash execution, web search, git operations, context providers, IDE integration.
And they send all of them to the LLM on every request.
Claude Code has 27 tools. Cursor has similar. Windsurf too.
When you say "fix the nav labels in pricing.html", here's what happens:
Edit
)That's like printing a 100-page menu when you're ordering a coffee.
One thing. It filters tool definitions.
Client sends 27 tools
↓
Proxy reads your message
↓
Proxy figures out what tools you need (6)
↓
Proxy forwards only those 6
↓
Provider responds normally
That's it. No magic. No LLM calls. Just regex matching and list filtering.
The proxy looks at your message and matches keywords:
"fix the nav labels in pricing.html"
regex "fix" → file_edit: +2
regex "change" → file_edit: +1
regex "html" → file_edit: +1
category = file_edit (score: 4)
Then it sends only the tools for that category:
KEEP: Read, Edit, Grep, Glob, Bash, Write
DROP: 21 others (GitHub, IDE, Context7, etc.)
WITHOUT proxy:
27 tools ████████████████████████████████ ~1,200 tokens
WITH proxy:
6 tools ██████████░░░░░░░░░░░░░░░░░░░░░░ ~300 tokens
Saved: ~900 tokens per request (75%)
Same model. Same quality. Just fewer schemas.
git clone https://github.com/dvcoolarun/dynamic-tool-proxy
cd dynamic-tool-proxy
pip install fastapi httpx uvicorn
python3 standalone_proxy.py
Then point your editor at it:
ANTHROPIC_BASE_URL=http://localhost:8090 claude
export ANTHROPIC_BASE_URL=http://localhost:8090
That's it. No API keys. No cloud accounts. No config files.
The proxy listens on port 8090. Your editor thinks it's talking to Anthropic. The proxy strips the tools and forwards the request.
The proxy classifies your message into one of 7 categories:
| Category | Tools Sent | When It Triggers |
|---|---|---|
| file_edit | Read, Edit, Grep, Glob, Bash, Write | "fix", "change", "edit", file extensions |
| search | Read, Grep, Glob, ListDirectory | "find", "search", "where is" |
| git | Bash, Read, Grep, Glob | "commit", "push", "branch", "git" |
| debug | Read, Grep, Bash, Glob | "error", "bug", "why", "broken" |
| web | WebFetch, WebSearch, Read | "url", "http", "website" |
| arch | Read, Grep, Glob, Bash, Agent | "refactor", "structure", "design" |
| full | All tools | Ambiguous messages (fallback) |
Ambiguous messages get all tools. No quality loss. Just in case.
I tracked it for a week. Here's what I saw:
| Metric | Before | After | Savings |
|---|---|---|---|
| Tokens per request | 1,200 | 300 | 75% |
| Requests per day | 100 | 100 | — |
| Tokens saved per day | — | 90,000 | — |
| Monthly cost saved | — | — | $150 |
At 100 requests per day, that's 2.7 million tokens per month you're not paying for.
Clients: Claude Code, Cursor, Windsurf, Cline, OpenCode, anything that uses Anthropic format.
Providers: Anthropic API, AWS Bedrock, Google Vertex, OpenAI-compatible, anything that accepts Anthropic format.
You don't need a specific provider. Point it at whatever you're already using.
RTK (if you're using it) solves a different problem:
| Dynamic Tool Proxy | RTK | |
|---|---|---|
| What it filters | Tool definitions (schemas) | Tool outputs (results) |
| When it filters | Before request sends | After tool runs |
| Savings | ~12k tokens per request | Varies by output size |
They're complementary. Use both for maximum savings.
1. Most developers don't check their token usage
They see the bill and accept it. They don't realize they're paying to send unused schemas to the model.
2. The biggest win is boring
It's just filtering a list. No LLM calls. No complex logic. Just "do you need this tool? no? remove it."
3. Local-first wins for developer tools
Developers want control. They want to see what's happening. They don't want another SaaS account with another API key.
4. Regex is enough
I thought I'd need LLM classification. I don't. Regex catches 90% of cases. Ambiguous messages fall back to full tool set. No quality loss.
GitHub: https://github.com/dvcoolarun/dynamic-tool-proxy
MIT license. Free forever. Your data stays local.
Star it if this saves you money.
Drop a comment. I'll answer every one.
Edit: Added streaming support. The proxy now passes through SSE streams unchanged.