Why OneTool Exists¶
MCP doesn't scale.
Here's the problem: every MCP server you add makes your AI dumber. Not in theory - in practice.
What's actually happening¶
Context rot is real¶
LLM performance degrades as input tokens pile up. Chroma's research measured this: the 10,000th token is handled less reliably than the 100th.
"Context must be treated as a finite resource with diminishing marginal returns." - Anthropic Engineering
MCP eats your context for breakfast¶
Every MCP server you connect adds thousands of tokens before you've said anything.
| Setup | Token Cost | What that means |
|---|---|---|
| 5 servers | ~55K | Gone before you type |
| 10+ servers | 100K+ | Context nearly full |
| Tool calls | N loops | LLM deliberation tax |
Each server makes context rot worse, not just bigger.
Vibe coding catches up with you¶
Collins Dictionary's 2025 Word of the Year: Andrej Karpathy's term for accepting code "that looks roughly right."
A UK study of 120 firms found 41% more debugging time at scale. Day 2 maintenance of AI-generated code is still an unsolved problem.
What OneTool does differently¶
One tool. Code execution. That's basically it.
The numbers¶
| Metric | Traditional MCP | OneTool |
|---|---|---|
| Token usage | 150,000 | 2,000 |
| Cost per query | $0.025 | $0.002 |
| Tool calls | 5+ | 1 |
98.7% fewer tokens. 10x cheaper. One call instead of five.
How it works¶
Traditional MCP:
Load tools (55K) → Reason → Call tool → Reason → Return
Total: ~150K tokens, 5+ reasoning loops
OneTool:
run request → Execute Python → Return
Total: ~2K tokens, 1 call
Instead of loading 50 tool schemas, LLMs write Python:
The approach¶
LLMs write Python to call functions. No JSON schema parsing. No tool selection loops. You see exactly what's being called because you wrote it.
15 packs built-in. Adding your own is dropping a Python file in a folder.
MIT Technology Review calls this shift "from vibe coding to agentic engineering." Whatever you call it, the token math works out.
What you get¶
- 24x fewer tokens, same accuracy
- Drop a Python file, get a new pack
- AST validation before execution
ot-benchfor testing LLM + MCP combinations- Proxy mode wraps existing MCP servers