Hand bounded work to a local model

@sezzlee/llm-mcp lets a planning agent, such as Codex or Claude Code, hand bounded language work to a model on your own Ollama. The agent passes a file path; the server reads the file, and only the model's short answer enters the agent's context.

You need Node.js 22 or later, jq and Ollama running.

1. Pull a model

Any Ollama model that can follow a JSON schema works. qwen3:8b is a good start on a machine with 8 GB of GPU memory:

ollama pull qwen3:8b

2. Pick a folder

The server reads files only inside its workspace, and writes only into an output folder inside it. To follow along, create one and save meeting-notes.txt, tickets.csv and operations-2026.txt into it:

mkdir -p ~/sezzlee-llm

3. Add the server to your client

Pass the workspace as an absolute path and name the model. Only the model is required; a client passes the server none of its own environment, so every setting goes in env.

claude mcp add local -e SEZZLEE_LLM_MODEL=qwen3:8b -e SEZZLEE_LLM_ROOT=/Users/you/sezzlee-llm -- npx -y @sezzlee/llm-mcp

Codex goes in ~/.codex/config.toml, Claude Desktop in claude_desktop_config.json (Settings → Developer → Edit Config, then restart), Cursor in .cursor/mcp.json and VS Code in .vscode/mcp.json. Without SEZZLEE_LLM_ROOT, the workspace is the folder the client starts the server in.

The Codex timeout covers the wait behind other calls and a long labelling run. These configurations were not loaded into the clients for this page: the Claude Code command was checked against claude mcp add --help, the Codex block follows the configuration a Codex-driving host generates, and the JSON follows each client's documented format.

4. Hand over a first task

Ask the agent: Summarise meeting-notes.txt for the team lead.

Expect one local_task call with kind set to summarize and the file's path. The server reads the notes and the local model summarises them. The agent receives the summary and nothing else.

See what the agent receives

The recipes call the tools the way an agent does, through the MCP Inspector, so you can see each answer exactly. Install the server once:

npm install -g @sezzlee/llm-mcp

Name the model you pulled:

export SEZZLEE_LLM_MODEL=qwen3:8b

Then define this helper in your shell. The Inspector passes the server none of your environment, so the helper hands over the model and the workspace with -e:

llm() {
  npx -y @modelcontextprotocol/inspector --cli sezzlee-llm -- \
    -e SEZZLEE_LLM_MODEL="$SEZZLEE_LLM_MODEL" -e SEZZLEE_LLM_ROOT="$HOME/sezzlee-llm" \
    --method tools/call --tool-name "$@" | jq '.content[0].text | fromjson'
}

This checks the model:

llm local_status | jq '{model, reachable, contextTokens, inputBudgetTokens}'

{
  "model": "qwen3.8:latest",
  "reachable": true,
  "contextTokens": 16384,
  "inputBudgetTokens": 7372
}

inputBudgetTokens is the most input one call may send: 45% of the 16,384-token window the server asks Ollama for, leaving room for the instructions and the answer. The answers in these docs were produced with qwen3.8:latest, a 27-billion-parameter Qwen model, so yours will read differently. Their shape is the same whichever model you use.

What it gives the agent

• local_task runs one task on the local model: summarize, extract, classify, transform or free, over short text or up to eight files. With a JSON Schema the answer is JSON.
• local_map labels every row of a CSV file with one of your labels and writes a labelled copy. The agent sees counts and a few sample rows, never the file.
• local_status says whether the model host answers and how much input one call may carry.

The local model gets one instruction and one input and cannot call tools. The server never cuts an input to fit: one the window cannot hold is split or refused.