# Matinta

Under construction.

Matinta is a small typed programming language with AI agents built into the language. I made it to be written by people and by any language model. An agent can use a local model (embedded llama.cpp), Claude or Ollama, and every answer is checked against a type. I created Matinta in October 2026, in Rust, to write the scenes of the Setealém engine, and the code is not published yet.

## Fact sheet

- Platforms: macOS
- Started on: 2026-10-30
- Built with: Rust, llama.cpp, whisper.cpp, Claude, Ollama
- Author: Antonio Leandro (https://linkedin.com/in/tonho)

## What does Matinta do?

### Agents in the language itself

agent, tool and ask are part of the syntax. Every question declares its answer type and a required fallback, and can have a deadline. Each agent has a token budget and only uses the tools it was given.

### The compiler teaches

Every error comes with a code, a line and the fix, also as JSON. Writing def, null, try or elif from another language gets the right form back. With only the guide, Claude Haiku 5.5 solved all 10 test tasks on the first try.

### Tests that repeat exactly

Tests live in the same file and run on a virtual clock with a test brain, with no waiting and no model calls. A law runs 100 cases, edge cases first, and shows the counterexample when it fails. --record and --replay repeat a run with real models, with no network.

### The right shape even from a small model

With a local model, Matinta turns the answer type into a grammar, so the model can only write what fits it. On 12 test questions, the embedded Qwen3 0.6B answered all 12 in the right shape. The same model through Ollama, without the grammar, got 9.

### Voice on the Mac itself

listen records the microphone, transcribe turns it into text with whisper.cpp and speak answers with a macOS voice. In one test, a recorded sentence became text and then a typed decision on Qwen3 8B, all local, in 12 seconds. An ask can also take an image.

### Every tool becomes an MCP tool

matinta mcp file.mt starts an MCP server where every tool in the file becomes a tool for Claude Code. The /// comment becomes the description and the parameters become the schema, with no JSON to write.

## In numbers

- 10/10 test tasks solved on the first try by Claude Haiku 5.5, reading only the guide (2026-10-30)
- 10/10 tasks solved by a local Qwen3 8B, with one chance to fix (2026-10-30)
- 12/12 answers in the right shape from the embedded Qwen3 0.6B (9 of 12 without the grammar) (2026-10-30)
- 2,100 tokens, roughly, in the guide the model reads (2026-10-30)

Let's talk: tonhowtf@gmail.com