On 17 August, in the middle of an 8-hour GitHub outage, Cursor launched its GitHub rival.

Origin is a git host built for agents, and the first product out of newly SpaceX-owned Cursor, which closed its $60B buyout 3 days earlier. 

What caught my eye: Origin ships without an MCP server, so while Cursor's own agent drives it natively, pointing Claude Code or any other MCP host at it turns up nothing to call.

That's what we're closing in this edition. About 120 lines of Node wrap the Origin CLI as an MCP server, so by the end Claude Code drives Origin on its own.

But first:

Huzzah: An experimental AI coding editor where you shape software with persistent, declarative pseudocode instead of firing longform prompts at an agent.

Shoehorn: A CLI that quantises any GGUF model down to the exact size that fits your free VRAM, then launches it with llama.cpp.

Visimer: An open-source WYSIWYG editor for Mermaid diagrams that lets you rename, reshape and rewire nodes by clicking the rendered diagram instead of editing syntax.

PyScrappy: An adaptive Python scraping toolkit and MCP server with self-healing selectors and TLS-fingerprint stealth that hands agents clean, LLM-ready data from any URL.

jit: A free, local-first macOS tool that finds plaintext secrets on your machine and tucks them behind Touch ID, injected just in time so your CLIs and AI agents keep working.

The server exposes 5 Origin subcommands as MCP tools over stdio: repo create, list, view and clone, plus auth status.

It boots clean, handles a missing or unauthenticated origin binary without falling over, and I ran it end to end before writing this up.

This puts the binary at ~/.local/bin/origin. If your shell says command not found: origin, add that to PATH:

Verify and sign in:

This also sets up the git credential helper, so git push and git pull against Origin remotes work without extra setup.

Sanity check:

Set package.json to ESM so import syntax works:

server.js:

Errors surface as tool output instead of crashing the server, because origin writes its failure reasons to stderr (expired auth, a repo that already exists), and you want those visible to the agent.

Test against the SDK's own client first, so that if something breaks once it's inside a host, you already know the server itself is clean and the problem is config.

test_client.mjs:

If it prints the 5 registered tool names and a "Logged in as ..." line, the server is clean and you can wire it into a host.

Add it as a project or user-scoped MCP server:

Or drop it straight into .claude/settings.json / .mcp.json:

Restart Claude Code and check the tool list picked up the 5 origin_* tools, then ask: "create an Origin repo called scratch-test and confirm it exists."

You'll watch it chain origin_repo_create into origin_repo_view on its own.

A couple of limits. 

Tool output comes back as plain stdout text unless you pass --json through origin_repo_view, which is fine for an agent reading it back, though anything downstream that expects a schema will struggle with it. 

The server also inherits whatever origin session is already on the machine, with no per-call token scoping, so if the agent can call origin_repo_create, it can create repos anywhere your account can write.

Put a permission gate in front of it before you ever run it unattended.

The thing I keep coming back to with MCP is how little the wrapper needs to understand.

My server has no idea what Origin does.

It shells out to the CLI and reads back whatever comes out, and Claude Code is just as much in the dark, yet once the tool schema is right the agent chains calls as if it's talking to a proper backend.

That's the part worth stealing.

Don't wait for full API access or credentials, write the MCP plumbing against whatever thin interface exists today, and swap in the backend once you're through the door.

The interface is where the work is, and what sits behind it can change whenever you like.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Speak your PR description, bug reproduction, or Cursor prompt. Wispr Flow auto-tags file names, preserves variable names, and formats everything for immediate paste into GitHub, Jira, or your editor.

No re-typing. No context gaps. No mangled syntax. Works natively inside Cursor, Warp, and every IDE at the system level.

4x faster than typing. 89% of messages sent with zero edits. Used by engineering teams at OpenAI, Vercel, and Clay.

Try Wispr Flow free

#AD 2