TypeSafe Mario

typesafe-computer-use is an open-source macOS computer-use tool that drives a Mac toward a plain-English goal. It captures the frontmost window, reads text with OCR, collects labelled controls from the macOS accessibility tree, parses dates, and sends a structured list of possible actions to the TypeSafe decision model instead of sending screenshots to a frontier model for every step. The loop chooses actions such as clicking, scrolling, browser navigation, waiting, or pressing keys; a writing model is called only for free-text entry, proposed URLs, and the final answer, while passwords are never typed. It supports dry runs, confidence and step limits, annotated captures, detailed action and probability logs, and offline replay of saved screenshots. The repository targets macOS 14 or newer with Python 3.12 or newer, requires a TypeSafe API key, and is licensed under the MIT License.

NanoJev is an open-source 0.6B parallel-decision model and training pipeline based on Qwen3-0.6B. It accepts states, questions, and candidate sets in a single forward pass and returns complete probability distributions without output-token decoding. Shared decision heads support dynamic-choice distributions over supplied candidates, Boolean proposition probabilities, and ordered score levels with expected values. The repository includes data generation, distribution-loss training, evaluation, persistent serving, and browser replay tools. Its game demonstrations combine model judgments with controller code: maze controllers ask local safety questions and track collisions and explored paths, while Snake controllers filter immediate collisions, plan toward visible food, and use the model to break ties between remaining actions. Checkpoints and datasets are distributed through Hugging Face, and the service exposes repeated batched evaluations through a local HTTP endpoint.

Mobile Jev is a standalone Android agent from droidrun that turns a written goal into device actions through the Mobilerun API, using TypeSafe's Jev to choose operations such as opening apps, tapping, entering text, scrolling, navigating, waiting, completing, or blocking. It observes the device, offers indexed controls and installed apps to Jev, validates the selected target against fresh screen data, executes the action, and observes the device again; it does not require an ADB connection. The executor resolves coordinates from observed bounds, rejects stale targets, records actions and execution traces, and verifies field values after text entry rather than relying solely on a model completion response. The project includes a local React studio with a live device stream, goal input, action timeline, task clock, model-latency measurements, stop and reconnect controls, and recent runs, along with a CLI for previewing or executing goals, inspecting devices, taking screenshots, profiling API timings, and issuing direct controls. It requires Node.js, pnpm, a Mobilerun Android device, and Mobilerun and TypeSafe API keys; it is intended for a single local operator, with recent runs held in memory. The repository is MIT licensed.

kev is an open-source, local Jev-inspired decision model from Jared Palmer. It runs on a laptop and accepts a document or other state plus typed questions, returning calibrated probabilities rather than generated text. It supports yes/no (noul), multiple-choice, and ordered-score questions. The model uses a LoRA adapter and a readout head on a Qwen base model. It packs the state and all questions into one sequence, applies a block-causal mask so each question can see the state but not sibling questions, and performs one prefill pass without decoding. A pointer head scores each option against the question's decision token and applies softmax; the head is trained with cross-entropy on labelled outcomes. The local server exposes a TypeSafe System One-compatible API at POST /v1/systemone, allowing the official TypeSafe SDK to connect by changing its base URL. It also provides model information, option-order permutation, and separate-pass comparison endpoints, plus a local playground. The repository includes 0.5B, 0.6B, 4B, and 8B model checkpoints and recipes for local or Modal-based training and evaluation. Serving is tested on Apple Silicon, with larger checkpoints requiring bf16 on a 32 GB Mac. The server has no authentication and is intended for local use. The project is licensed under Apache-2.0; base models and datasets have separate licenses.

SuperAstra is a SNES-themed desktop companion that uses natural-language prompts to investigate and alter a running game through BizHawk. Its agent receives screenshots, cartridge identity data, emulator registers, memory scans, hardware-domain reads, controller probes, and checkpoint results; it can search cartridge data, identify game-memory structures, create routines or guarded patches, test changes, and retain cartridge-specific findings in a local notebook. Each committed mutation can be undone through emulator checkpoints, while named experiment checkpoints support controlled trials that restore the player's prior state. The prototype runs on Windows or Linux with Python and Tkinter, BizHawk using its BSNES SNES core, a ROM, and an OpenAI API key configured for the stated model; it also includes limited local Mario shortcuts that do not require an API key. It does not patch the original ROM file, create exportable ROM patches, expand ROM capacity, or guarantee success on unfamiliar games. The original source is MIT-licensed.