OneTool MCP

🧿 One MCP for developers - no tool tax, no context rot. 250+ tools your agent calls as Python code: search, docs, files, databases, diagrams, vision, memory - plus a proxy for every MCP server you already use.

Works with Claude Code, Cursor, Codex - any MCP client

---

The Problem

Every MCP server re-sends its tool definitions on every request: 3K-30K tokens each. Connect 5 servers and you've burned 55K tokens before the conversation starts. Connect 10+ and you're at 100K.

The math is brutal: Claude Opus 4.5 at $5/M input tokens, 20 days × 10 conversations × 10 messages × 3K tokens = $30/month per MCP server - even if you never use the tools.

And then there's context rot - your AI literally gets dumber as you add more tools (Chroma Research, 2025).

The Solution

OneTool is one MCP server that exposes tools as a Python API. Instead of reading tool definitions, your agent writes code:

__onetool brave.search(query="react 19 server components")

Configure one MCP server. Use unlimited tools - ~2K tokens no matter how many you add.

"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

97% fewer tokens. 30× lower cost. No context rot. (Measured - 47,660 → 1,131 input tokens against 18 MCP servers.)

📖 Read the full story

---

Code Is the Interface

Because tools are Python functions, your agent does things tool-call JSON can't: batch, chain, loop, compose.

__onetool

page = webfetch.fetch(url="https://fastmcp.dev/changelog", output_format="markdown") notes = ot_llm.transform(data=page, prompt="Summarise the breaking changes") mem.write(topic="deps/fastmcp", content=notes)

Three packs, one request. Intermediate results flow between tools as variables - the page body never touches your context window, and the summarising runs on a cheap model instead of your expensive coding agent.

Every call is explicit and reviewable - __onetool brave.search(query="...") shows you exactly what runs. No tool-selection guessing.

And the runtime is built for how agents actually type:

• mem.search(q="auth") works - any unambiguous parameter prefix resolves (q= → query=)
• wb.draw(...) works - packs have short aliases (wb, ctx, img)
• github.listRepositories() works on proxied servers - snake/camel/Pascal all resolve
• A typo'd tool gets a did-you-mean, a disconnected server names the command that fixes it
• Oversized results come back as a searchable handle instead of flooding the window

---

Install

Bootstrap (installs uv if missing, installs OneTool, initialises config, prints MCP config):

curl -LsSf https://onetool.beycom.online/install.sh | sh          # macOS / Linux

irm https://onetool.beycom.online/install.ps1 | iex # Windows (PowerShell)

Or install manually with uv:

uv tool install 'onetool-mcp[all]'   # everything

onetool init --config ~/.onetool

Then print ready-to-paste MCP client config with resolved absolute paths and add it to your client (claude-code, claude-desktop, cursor, or vscode):

onetool init mcp-config --client claude-code   # or omit --client for all four

That's it. All 250+ tools work out of the box.

Verify: onetool init validate --config ~/.onetool/onetool.yaml

Install the ot-ref skill into your agent with vercel-labs/skills - it teaches the call conventions and ships a greppable index of every tool signature:

npx skills add https://github.com/beycom/onetool-mcp --skill ot-ref --agent claude

📖 Full installation guide

---

What's Inside

---

Tools

28 packs, 253 tools ready to use (console in beta):

📖 Complete tools reference — every signature, generated from source

---

MCP Server Proxy

Keep the MCP servers you already use. Wrap them in YAML and call them explicitly - as Python namespaces, without their tool tax:

# .onetool/onetool.yaml

servers: local_tools: type: stdio command: npx args: ["-y", "some-mcp-server@latest"] private_api: type: http url: ${PRIVATE_MCP_URL} auth: type: bearer token: ${PRIVATE_MCP_TOKEN}

__onetool private_api.read_resource(path="README.md")

Proxied servers can be enabled, disabled, and restarted mid-conversation with ot_servers - no client restart.

📖 Configuration guide

---

Secrets You Can Commit

onetool init walks you through encrypted secrets: values in secrets.yaml are age-encrypted, the private key lives in your OS keychain, and decryption happens transparently at load.

# secrets.yaml - safe to inspect, safe to commit

brave_api_key: age1enc:YWdlLWVuY3J5cHRpb24ub3JnL3YxCi0+IFgyNT...

📖 Security guide

---

Use from the CLI

Works as an MCP server and as a direct CLI bridge into the same running process - loaded config, secrets, and proxy connections stay warm. Useful for agent harnesses, scripts, and automation:

# Recommended local MCP root mode: stdio

onetool serve --config .onetool/onetool.yaml

URL-based MCP root mode for containerized clients

onetool serve --transport http --config .onetool/onetool.yaml --host 127.0.0.1 --port 8767 --path /mcp

Enable the MCP-owned direct API in onetool.yaml:

direct.host.enabled: true

Start OneTool as MCP, then use the port printed in startup logs.

onetool