A Neovim plugin is provided in extra/lectic.nvim that offers support for the .lec filetype, including syntax highlighting, streaming responses, folding of tool use blocks, keymaps, and commands for updating conversations, consolidating memories, and elaborating model responses. For more details, check out the README

A VSCode plugin is provided in extra/lectic.vscode that offers support for the .lec filetype, including syntax highlighting, streaming responses, folding of tool use blocks, keymaps, and commands for updating conversations, consolidating memories, and elaborating model responses. For more details, check out the README

Most text editors support piping through external commands, like linters and code formatting tools. You can use lectic in the same way you would use one of those tools: by sending the buffer to lectic on stdin, and replacing the buffer contents with the result from stdout.

The exec tool allows the LLM to execute commands directly. For security, you can configure which commands are available and optionally run them in a sandbox.

tools:
    - exec: python3             # Allow LLM to run Python
      name: python              # Optional custom name
      usage: "Usage guide..."   # Instructions for the LLM
      sandbox: ./sandbox.sh     # Optional sandboxing script
      confirm: ./confirm.sh     # Optional confirmation script

Example conversation using the exec tool:

Could you check if this code works?

:::Assistant

I'll test it using Python:

{python}
print("Hello, world!")
{/python}

The code executed successfully and output: Hello, world!

:::

[!WARNING] If you provide your LLM access to any dangerous commands, you should take some safety precautions.

When an exec tool is configured with a sandbox script, the command and its arguments are passed to the sandbox script which is responsible for executing them in a controlled environment. For example, the provided extra/sandbox/bwrap-sandbox.sh uses bubblewrap to create a basic sandbox with a temporary home directory.

When an exec tool is configured with a confirm script, the confirm script will be executed before every call to that tool, with two arguments: the name of the tool, and a JSON string representing the arguments to the call. If the confirm script returns a nonzero exit status, the tool call is cancelled. An example confirmation script, using zenity is included in this repository at extra/confirm/zenity-confirm.sh.

The sqlite tool gives the LLM the ability to query SQLite databases. You can configure limits and provide schema documentation. The schema is automatically supplied to the LLM.

tools:
    - sqlite: ./data.db         # Path to SQLite database
      name: query               # Optional custom name
      limit: 10000              # Maximum size of serialized response
      details: this contains... # Extra details about the DB

Example conversation using the sqlite tool:

What are the top 5 orders by value?

:::Assistant

I'll query the orders table:

{query}
SELECT 
    order_id,
    total_value
FROM orders
ORDER BY total_value DESC
LIMIT 5;
{/query}

Here are the results showing the largest orders...

:::

This tool lets you give your LLM an opportunity to deliberately pause and think about something, in the style suggested by anthropic's engineering blog. It provides the LLM with some "scratch space" to think a little bit before speaking "out loud".

tools:
    - think_about: what the user is really asking   # A thing to think about
      name: thinker                                 # Optional custom name
      usage: Use this whenever user is imprecise.   # Optional extra usage advice

Example conversation using the think tool:

So what's the best place in Boston?

:::Assistant

Hmmm....

{think}
What are they really asking? What might "best" mean? That probably depends on
what kind of place they're interested in. Do they want to know about the best
places to visit, to live, to work? I had better ask.
{/think}

What kind of place are you interested in? Places to live, places to work,
places to visit, or something else?

:::

This tool lets you give your LLM access to a simple web server, so that it can serve you up web pages and apps that it generates, a bit like Claude's artifacts. and analysis tool. The web page should automatically open in your browser (the conversation will be blocked until the page is loaded). Once the page is loaded, the server will be shut down.

tools:
    - serve_on_port: 9000   # serve on localhost:9000
      name: my_server       # Optional custom name

Example conversation using the think tool:

Could you serve me up a little tic-tac-toe game?

:::Assistant


{server}
A WHOLE BUNCH OF HTML, JS, and CSS
{/server}

There you go! Your tic-tac-toe game is available on localhost:9000

:::

Lectic lets you use model context protocol servers to provide tools for your LLMs. There are already a huge number of servers available: check out this list provided by the MCP organization, and also awesome MCP servers. To provide access to an MCP server, you can add the server as a tool like this:

tools:
    - mcp_command: npx # The main command to launch the server
      name: brave      # Optional name, used in resource content references (see above)
      roots:           # Optional list of filesystem roots
        - /home/graham/research-docs/
      args:            # Additional arguments to the main command
        - "-y"
        - "@modelcontextprotocol/server-brave-search"
      env:             # Environment for the server to run in.
        - BRAVE_API_KEY: YOUR_KEY_GOES_HERE
      confirm: ./confirm.sh # Optional confirmation script
      sandbox: ./sandbox.sh # Optional sandboxing script
    - mcp_sse: URL_GOES_HERE # URL for a remote MCP server that uses SSE
      confirm: ./confirm.sh
    - mcp_ws: URL_GOES_HERE # URL to a remote MCP server that uses Websockets
      confirm: ./confirm.sh

Roots are supported, and giving a server a name lets you retrieve any resources that the server provides using content references. The LLM will also be provided with a tool that allows it to list resources that the server makes available.

Example conversation using the MCP tool:

Could you do a brave web search for cool facts about ants?

:::Assistant

Sure!

{brave_search_tool}
query: cool ant facts
{/brave_search_tool}

Did you know that ants have two stomachs, one for themselves and one for food 
they're going to share with the colony?

:::

[!WARNING] By default, lectic grants LLMs full access to tools provided by MCP servers. If you have a server that provides potentially dangerous tools, you should take some safety precautions.

Lectic's safety mechanisms are intended to protect you from an LLM making mistakes. They offer only very limited defense against a genuinely malicious MCP server. There are serious security risks associated with the MCP protocol. You should never connect to an untrusted server. See this post, for examples and details.

When an local MCP tool is configured with a sandbox script, the command that starts the MCP server, along with its arguments are passed to the sandbox script which is responsible for executing them in a controlled environment. For example, the provided extra/sandbox/bwrap-sandbox.sh uses bubblewrap to create a basic sandbox with a temporary home directory.

When an MCP tool is configured with a confirm script, the confirm script will be executed before every call to that tool, with two arguments: the name of the tool, and a JSON string representing the arguments to the call. If the confirm script returns a nonzero exit status, the tool call is cancelled. An example confirmation script, using zenity is included in this repository at extra/confirm/zenity-confirm.sh.

Native tools give you access to some of the built-in functionality that certain LLM backends provide, for example built in search or code execution environments. At the moment two kinds of native tools are supported: search and code. Native search lets the LLM perform web searches, and native code lets it execute code in a remote sandbox for data analysis tasks. You can provide access to native tools like this:

tools:
    - native: search # provide a native search tool
    - native: code   # provide a native code sandbox tool

Native tool support varies by LLM provider. Right now, Lectic supports native tool use with gemini and anthropic.

• Recent gemini models support both search and code. But they're subject to some limitations imposed by the Gemini API. You cannot provide more than one native tool at a time, and you cannot combine native tools with other tools. (If you try, the API will throw an error. If you find that it doesn't, Google must have lifted this limitation—in that case, let me know!)
• Anthropic support search only. For more information, you can read this announcement.