Working from your own editor

The dashboard's AI Builder is one way to write a tool. The other is your own editor — VS Code, Cursor, Claude Code, whatever you already use — with its AI agent working on the tool's folder. Until now that meant a round trip: Download Code, edit, Replace Code. Voshi's MCP server removes the round trip. Point your editor's agent at it once, sign in, and the agent can read your tools, fetch their code, upload what it changed, and publish versions, from inside the conversation you are already having with it.

What an MCP is

MCP (Model Context Protocol) is the standard way an AI agent is given tools to call. An editor's agent — Claude Code, Cursor's agent, VS Code's Copilot agent — can talk to any MCP server you configure, and from then on the server's tools are things the agent can do, the way reading a file or running a command are. Voshi runs one such server for builders at:

https://api.link.voshi.com/ltiaas/mcp

Connecting to it does not change what a hosted tool is: the same folder, the same manifest.json, the same rules in AGENTS.md. It changes who carries the folder back and forth.

Why connect

• Edit in place. Ask your agent to "add a timer to the quiz," and it reads the tool's current files from Voshi, edits them, uploads the result as the draft, and tells you to refresh the Sandbox. No zip, no folder picker.
• Work on several tools. The agent sees every tool you are a member of and switches between them by name.
• Publish from the editor. When the Sandbox looks right, "publish this as a private version" is a sentence, not a trip to the dashboard.
• Keep your own tooling. Git, your linter, your tests, your preferred model — the tool's code lives in your folder like any other project, and Voshi is where it is deployed.

The AI Builder is still there, and both work on the same draft: a build in the dashboard replaces what your agent uploaded, and the other way round. Use whichever suits the moment.

What the agent can do

Once connected, the agent has these tools. It picks them itself from what you ask; you do not call them by name.

Tool	What it does
read_contract	Reads the tool's AGENTS.md: the rules of a hosted tool, and the tool's locations. Agents call it first.
list_apps	The tools you are a member of, with their current and pilot versions.
get_app	One tool in detail: its locations, its draft, and the dashboard and Sandbox links.
read_files	The draft's file list, and the text of the files the agent asks for.
stage_upload, commit_upload	Upload a folder as the new draft. The agent zips the folder and posts it straight to Voshi's storage with a command the server gives it, then commits; the same checks run as for Replace Code, and a refusal names the file and the rule.
publish_version	Publish the draft as the next version, private by default, with a note.
list_versions	The tool's versions, with their status and visibility.

Two things an agent cannot do from the editor, on purpose:

• Launch the tool. Testing needs a browser and a real launch. get_app hands the agent the Sandbox link; you open it.
• Change the tool's locations. Those are rows on the dashboard's Locations tab, and AGENTS.md lists them so the agent handles every one; the agent reads them and codes for them, and you manage the list.

Security

• You sign in, the agent does not. Connecting opens a browser tab on Voshi where you sign in with your builder account and allow the connection. The agent never sees your password.
• The connection reaches only these tools. What the agent holds is not your dashboard login: it cannot sign in as you, read anything but your tools' code, or change your account. A leaked editor configuration exposes the tools you build and nothing else.
• It expires in two weeks, after which the agent asks you to allow it again. Nothing renews it silently.
• You can disconnect at any time from the dashboard's Editor Agents page, which lists every agent you have allowed, with when it expires.

note

What your agent writes is still your tool: the same upload checks apply, and a version it publishes goes through the same review before it reaches instructors. An agent cannot put code in front of students that you could not.

Ready? See Connecting your editor.