The flagship collection
13 projects, with different jobs and different levels of readiness. Pick the problem you have, then follow the guide to a first useful result.
Start with a small task in your own environment. The guides cover prerequisites, commands, and what to check.
Published CLI and MCP server
Find the files that matter before changing a repo.
Start with the packaged refresh-token example: map the code, locate its implementation and test, and load a focused context bundle without an API key. Then use the same commands with a task from your own repository.
For: Developers and coding agents working in an unfamiliar repository.
Published Python package
Turn a web page into Markdown your pipeline can use.
Fetch a public URL, preserve the content you need, and choose how much detail to pass downstream. Includes a CLI, Python API, MCP server, bounded crawling, and optional browser rendering.
For: Developers building research, retrieval, and web ingestion workflows.
Development preview
Keep useful project context across coding sessions.
Local memory for coding agents with separate project, workspace, and global scopes. Inspect what was recalled, update stale facts, and keep durable context in a local SQLite database.
For: Solo developers who want inspectable continuity between coding sessions.
Published MCP server
Let your agent ask you a question in Discord.
Connect an MCP host to a Discord bot so an agent can ask for clarification, wait for a reply, and continue with your answer. Threaded conversations keep each task together.
For: Developers who want to answer agent questions away from their editor.
Published protocol SDK
Give agent coordination a compact, explicit message format.
Encode bounded coordination intents with a shared vocabulary, parse them at the receiver, and fall back when an intent needs more context. The Python SDK makes the protocol available without loading a language model.
For: Engineers building a controlled message channel between cooperating agents.
Alpha / active rewrite
Find and compose an existing skill before creating another.
A skill acquisition workflow for coding agents: search installed and curated packages, compose what is available, then create a local draft when something is missing. New drafts remain untrusted until review.
For: Developers experimenting with reusable skills and reviewable agent self-extension.
Published checkpoints and training experiments. Check the model card and hardware requirements before downloading.
Open research checkpoints
Explore a small language model trained from scratch.
A 272M-parameter Mixture-of-Experts family with an instruction-tuned checkpoint and task specialists. Start with one named checkpoint, inspect its model card, and evaluate it on a narrow task.
For: Model researchers and developers evaluating small custom language models.
Experimental foundation checkpoint
Study a custom architecture through an open checkpoint.
The public 0.5B SABER checkpoint is a foundation-stage model for architecture experiments. Its files, custom modeling code, and training-stage notes are available for inspection.
For: Researchers comparing custom transformer components and training behavior.
Methods, prototypes, and competition work. Each guide distinguishes the artifact from what remains to be validated.
Published research library
Experiment with an attention cache that stays bounded.
A PyTorch attention module combining a recent-token window with exact and summary memory banks. Inspect fixed-size cache behavior on synthetic inputs before attempting model integration.
For: Inference engineers and researchers studying memory-limited attention.
Open research implementation
Test whether a repeated JSON schema deserves its own vocabulary.
A structure-aware tokenizer that learns JSON keys and value fragments. Train it on your schema, inspect the token sequence, and compare token counts and round-trip behavior on held-out examples.
For: Researchers building structured-data models or custom encoding pipelines.
Public competition experiments
Study the tradeoffs behind a language model under a 16MB limit.
Public submissions and engineering notes from OpenAI's Parameter Golf challenge: model architecture, quantization, training kernels, and evaluation under strict artifact and compute limits.
For: Engineers studying compact model training, quantization, and reproducible evaluation.
Research prototype
Test whether a rotation aligns two models' internal representations.
An experimental pipeline for extracting hidden states, fitting an orthogonal alignment, and comparing cross-model decoding against controls. The result is an experiment, including when the hypothesis fails.
For: Interpretability researchers studying representation alignment between language models.
Technical analysis and practical walkthroughs.
Published articles and newsletter
Follow the evidence behind the AI claims.
Technical walkthroughs, reporting, and analysis of the systems people are building with AI. The Checkpoint collects developments worth examining more closely.
For: Engineers and readers who want context, sources, and a technical argument they can examine.