MATLAB Local Agent Quickstart

Run a local AI coding agent against MATLAB®, with no code, data, or prompts sent off the machine.

This repository connects three pieces: a local LLM runtime, the OpenCode coding agent, and the MATLAB MCP Server with settings tuned for local models, then verifies the result with five demos.

Requirements

• MATLAB R2025a or later (R2025a enables the plain-text Live Script format one demo produces)
• Git™
• The MATLAB MCP Server
• One local LLM runtime: Ollama, LM Studio, or llama.cpp
• One agentic coding tool: OpenCode
• Hardware per the tier table below

Step 1: Prepare

Get the repo:

git clone https://github.com/matlab/matlab-local-agent-quickstart.git

Then run the readiness check in MATLAB:

cd matlab-local-agent-quickstart/prepare
checkLocalAI

It detects your memory, GPU, MATLAB version, Git, and Node.js®, assigns your machine a tier, and prints the recommended model, quantization, context window, and harness. It makes no network calls.

Tier	Hardware	Recommended model	Context	Harness	Skills
Small	16 GB Apple® silicon; 12 GB NVIDIA® GPU	Gemma 4 12B QAT, Q4	8k-16k	OpenCode	off
Medium	32 GB+ Apple silicon; 24 GB+ NVIDIA GPU	Qwen3-Coder-30B-A3B, Q4_K_M	16k-32k	OpenCode	curated

Small tier supports single-step assisted help. Medium should be used for reliable autonomous work.

Step 2: Set Up

Follow one guide per piece, in order:

1. MATLAB MCP Server: download the binary and note its path.
2. Runtime: Ollama · LM Studio · llama.cpp
3. Harness: OpenCode

Each guide applies a template from setup/configs/: the runtime setup sets the context window explicitly, and the OpenCode template connects to your local endpoint and registers the MCP Server with telemetry disabled.

Set the context window with care. The context window is how much text the model can hold at once. Your code, the conversation, and the agent's own working notes all share it. Local runtimes default to a small window, often far below what the model is capable of. When a session outgrows that window, the oldest content is dropped silently: the agent appears to "forget" earlier files or instructions, with no error to signal why.

Then verify the whole stack from MATLAB, naming your runtime and your tier's context value:

cd setup
verifySetup                                            % Ollama, Medium tier (16384)
verifySetup(runtime="lmstudio", contextTokens=12288)   % LM Studio, Small tier
verifySetup(runtime="llamacpp")                        % llama.cpp server

Verification confirms each link by name (runtime endpoint, model loaded, MCP Server binary if you pass its path) and prints the budget for the context window you configured: harness overhead, tool schemas, what remains for your code and conversation.

Step 3: Run the Demos

Five demos, each with starter assets, the exact prompt to give your agent, and a verification script that checks the outcome.

Demo	What happens	Minimum tier	The result you get
Fix the Error	Agent diagnoses and fixes a failing pipeline, then reruns it	Small	Error to green run, root cause named
Modernize the Codebase	Agent clears Code Analyzer warnings and proves behavior unchanged	Small-Medium	Warning count to zero, outputs verified equal
Test the Legacy Code	Agent writes and runs a test suite for untested utilities	Small-Medium	Passing suite, a real bug found, mutation check passed
Make It Fast	Agent vectorizes a slow script and measures the speedup properly	Medium	Speedup multiple, measured with timeit
Analyze the Sensor Data	Agent finds a bearing fault frequency in vibration data and writes a report	Medium	Detected frequency vs. physics-predicted value

Every demo runs on base MATLAB with no additional toolboxes. When you move from the demos to your own code, docs/starter-prompts.md shows the bounded-task prompt framing that keeps local models productive.

Troubleshooting

Below is a list of some of the failure modes you may encounter.

• Context truncation.
• Incorrect tool calls.
• Skills failing to trigger.
• Streaming parse errors.
• Chat-template mismatches.

Advice on fixing these issues is provided in docs/troubleshooting.md

Related Projects

• MATLAB MCP Server: the tools your agent uses to run, test, and analyze MATLAB code
• MATLAB Agentic Toolkit: curated skills for MATLAB workflows
• Simulink Agentic Toolkit: curated skills for Simulink® workflows
• MATLAB AI Agent SDK: create your own agents in MATLAB

License

The license is available in the LICENSE file in this GitHub repository.

Contact

To ask questions, report issues, or request technical support, open an Issue.

---

Copyright 2026 The MathWorks, Inc.

---