OneTool

🧿 One MCP for developers - No tool tax, no context rot.
100+ tools including Brave, Google, Context7, Excalidraw, AWS, Version Checker, Excel, File Ops, Database, Playwright, Chrome DevTools and many more.

Learn OneTool Reference

OneTool v2 is here

New in v2 — highlights:

Dynamic AWS proxy — the only practical way to use all 57+ official AWS MCP servers; zero token tax, credentials and SSO handled automatically
Live whiteboard — draw diagrams and architecture with a Mermaid-compatible DSL, powered by Excalidraw
Three search engines — Brave, Google (Ground), and Tavily; each with batch support, topic filters, and answer summaries
Browser annotations — highlight page elements and guide users through multi-step workflows via Chrome DevTools or Playwright
Interactive setup — onetool init opens a TUI to select exactly the extensions you need; no manual YAML editing to get started
Cleaner config — flat ~/.onetool/ layout, explicit --config and --secrets flags, and a versioned schema with clear errors
Encrypted secrets — age-encrypted secrets.yaml backed by your OS keychain
Leaner install — optional [util] and [dev] extras; install only the dependencies you need
Smart context store — ot_context (ctx); SQLite+FTS5 store; search and navigate large tool outputs without filling context
Image vision — ot_image (img); routes to a cheaper, better vision model; zero host tokens; supports local files, URLs, clipboard, and 8 formats (PNG, JPEG, GIF, WebP, TIFF, HEIC, AVIF, SVG)

See everything that's new in v2

The Problem¶

Tool Tax¶

Each MCP Server consumes between 3K and 30K (looking at you, GitHub MCP!) in tokens per request. Every single request. The maths is brutal: If you're using Claude Opus 4.5 at $5 per million input tokens - 20 days × 10 conversations × 10 messages × 3K tokens = 6M input tokens. You waste approx. $30 in Tool Tax per MCP Server, per month.

Context Rot¶

And then there's context rot - the phenomenon where the agent's performance degrades as the context window fills (Chroma Research, 2025). Every tool description pushes valuable conversation history out of the context window. Your AI literally gets dumber as you add more tools.

The Solution¶

OneTool is one MCP server that exposes tools as a Python API. Instead of reading tool definitions, your agent writes code - brave.search(q="react docs") - and OneTool runs it.

Configure one MCP Server. Use unlimited tools.

96% fewer tokens. 30× lower cost. No context rot. 100+ tools, extensible and configurable

Read the full story

Features¶

96% Token Savings¶

MCP servers consume 3-30K tokens before you start. OneTool uses ~2K tokens no matter how many tools and MCP servers you add. No tool tax. No context rot. 30× lower cost.

"Agents scale better by writing code to call tools instead. This reduces the token usage from 150,000 tokens to 2,000 tokens...a cost saving of 98.7%" - Anthropic Engineering

See comparison

Code, Not Tool Calls¶

Agents are excellent at writing code. OneTool tool calls can be batched, chained, and used just like any other Python function.

Learn more

Explicit Execution¶

Write Python, not tool definitions. You see exactly what runs. No more guessing.

>>> brave.search(q="AI")

Forge New Tools¶

Build new tools as part of the conversation. New tools are just Python functions. Drop a file, get a pack.

Create tools

MCP Server Proxy¶

Wrap any existing MCP server. Configure in YAML. Call explicitly - without the Tool Tax. Pre-configured: Chrome DevTools, GitHub.

Use other MCP servers

Great Developer Experience¶

Snippets with Jinja2. Aliases for common tools. Parameter prefixes (q for query). Single YAML config.

Configuration¶

Single, well-structured YAML config with global and project scopes. Per-pack settings for timeouts, limits, models.

Observability¶

Meta tools for introspection. Structured logging with LogSpan. Runtime statistics for costs and success rates.

100+ Tools¶

Web Search (Brave, Google), Context7, Version Check, Chrome DevTools, Playwright, Excel, File Ops, Database, AWS, Memory and many more.

Browse tools

Image Vision¶

Routes image analysis to a cheaper, better vision model. Zero tokens charged to your host session. Load from local files, URLs, or clipboard.

Supports PNG, JPEG, GIF, WebP, TIFF, HEIC, AVIF, SVG.

>>> img.load(img="invoice.png")
>>> img.ask(img="#inv_01", q="Extract all line items and prices")
>>> img.clip_ask(q="What does this screenshot show?")  # clipboard shortcut

ot_image reference

Smart Context Store¶

Persistent SQLite+FTS5 store for large tool outputs. Write once, search many times. Agents build up knowledge across tool calls without burning context window.

>>> ctx.write(content="...", source="research")
>>> ctx.search("async error handling")

ctx reference

Smart Tools¶

Delegate to cheaper agents. Fetch a page, summarise with Gemini Flash ($0.50/M), pass back to Opus ($5/M). 10× savings.

Smart Tools

Security¶

Multiple layers built in: isolated secrets.yaml, AST validation, path boundaries, output sanitisation.

Testing & Benchmarking¶

The bench harness compares tools against other MCP servers using a real agent. Define tasks in YAML. Get objective metrics: token counts, costs, accuracy scores, timing.

The Bench Tool

Engineering Practices¶

2,000+ tests (smoke, unit, integration). OpenSpec for change proposals - specs before code. Type hints throughout. Ruff + Mypy.