What MCP adds
The Model Context Protocol (MCP) is an open standard for connecting AI apps to tools and data. A program that offers tools this way is an MCP server. Laghu’s agent is an MCP client: connect a server, and three kinds of thing join the agent.
- Tools it can use, such as reading a repository’s history, fetching a web page or opening a GitHub issue.
- Documents a server offers to read, which the agent can list and open.
- Prompts, ready-made requests a server provides. They appear as buttons in the Agent screen and fill in the message box.
The model doing the work still runs on your own PC.
Add a server
Open Agent in the side bar and scroll to MCP servers. There are four ways in.
- From the ready-made list. Files, Fetch, Git, Memory, Time, Sequential thinking, Playwright and GitHub. Press Add, fill in the folder, repository or token it asks for, and it starts. Most of these run with Node.js or uv, which the card tells you if your PC lacks.
- A command on this PC. Give it a name, the command (for example
npxoruvx) and its arguments. - A web address. For a server on the web, give its address and, if it needs one, a token, sent as
Authorization: Bearer. - Paste a config. The JSON you already use in Claude Desktop, Cursor or VS Code works as it is.
A token can also be written as ${GITHUB_TOKEN}, and Laghu reads it from the environment, so it need not sit in the config file.
What Laghu supports
- Connections: servers on your PC (stdio), servers on the web (Streamable HTTP), and older web servers that use HTTP with server-sent events.
- Protocol versions: 2025-06-18, 2025-03-26 and 2024-11-05, agreed with each server when it connects.
- Tools: long tool lists over several pages, tools a server adds or removes while running, and every kind of result: text, structured data, documents and links. A picture a tool returns is shown to a model that can see.
- Documents and templates: listed and read by the agent.
- Prompts: listed in the Agent screen, with their questions asked before they fill in the message box.
- Questions from a server: when a server asks you something, the question appears in the conversation. Laghu also tells a server which folder the agent works in.
Laghu does not offer its model to servers to write with (the protocol calls this sampling).
Many tools, small models
Every tool switched on is described to the model on every step, and a model running on a laptop reads a limited amount at a time. So each server in the Agent screen shows its tools with a switch for each, and how much of every step they take. Switch off what you do not need.
When more tools are on than fit, Laghu sends the ones that match the task and gives the model a way to search for the others. A tool it finds can be used from the next step.
Staying safe
- Every MCP tool asks first. A server can describe its tools, and may say a tool only reads, but that is its own claim. Laghu shows it on the card and leaves the decision to you. You can allow a tool you trust for good.
- Tokens stay on your PC. They are kept in Laghu’s config file, or read from the environment.
- Know where requests go. A server on your PC can do what any program can. A server on the web receives the requests the agent sends it. Add servers you trust.
Questions
Does Laghu work with MCP servers?
Yes. Its agent connects to MCP servers on your PC or on the web, and uses their tools, documents and prompts.
Can I use the MCP config from Claude Desktop?
Yes. Paste it into Laghu’s Agent screen under MCP servers, and the servers in it are added and started.
Do I need Node.js?
Only for servers that run with npx. Servers that run with uvx need uv, and servers on the web need neither.
Which model works best with MCP tools?
A model badged Agent in the library, such as Qwen2.5 7B or Qwen3 8B. Switching off tools you do not need helps any model.