AI & ML
Models, agents, inference engines, training and evaluation code.
12 projects
OpenBot gives every AI coworker its own browser and an audit trail
CopilotKit/OpenBot
CopilotKit's self-hosted template runs each agent in its own container behind a policy gateway. Out of the box, that policy allows everything.
TrueForge is TrueFoundry's open agent harness, priced against Claude
truefoundry/trueforge
An MIT-licensed TypeScript runtime for agents: model calls, MCP tools, skills, sandbox and approvals. Its cost benchmark is 14 tasks, run by the vendor.
FreeToken streams MoE experts to a gaming GPU on demand
FlashML-org/FreeToken
A Python and CUDA engine that keeps hot experts on the GPU, the rest in RAM, and splits misses between PCIe and the CPU. One outside test: 2.25x Ollama.
Grok Build is SpaceXAI's coding agent, open in Rust
xai-org/grok-build
A full-screen terminal agent with headless mode and an editor protocol. The repository is a periodic sync out of a closed monorepo, not the monorepo itself.
Slate signs every invoice PDF and keeps its AI to drafts
slatecoil/app
Israeli invoicing that signs each document with Ed25519, gives it a public verification page, and keeps the AI to drafts a human approves.
ADHD fixes premature convergence by splitting the context
UditAkhourii/adhd
A skill that spawns parallel reasoning processes under distorted cognitive frames with no shared context, then scores, prunes traps and deepens survivors.
Colibrì runs 744B to 2.8T models off your SSD, in C
JustVugg/colibri
An engine treating VRAM, RAM and disk as one hierarchy, with a hard rule: run out of fast memory and it slows down, it never silently changes the model.
OpenScience gives a research agent 295 skills and a lab
synthetic-sciences/openscience
A browser workbench that reads the literature, writes code, runs experiments and queries UniProt, PDB, ChEMBL and thirty more databases, on your own API keys.
TurboFieldfare runs Gemma 4 26B on an 8 GB MacBook
drumih/turbo-fieldfare
A Swift and Metal runtime that keeps a 1.35 GB core resident and streams only the experts each token needs off SSD, built for Apple Silicon rather than on MLX.
A 2.78-trillion-parameter model in 8 GB of RAM, in C99
FareedKhan-dev/kimi-k3-in-c
kimi-k3-in-c streams a 1.56 TB checkpoint off disk to run Kimi K3 on an ordinary laptop CPU. No BLAS, no framework, no GPU, 176 KB of engine.
QM is an agent with a per-employee sandbox, in Slack
yc-software/qm
A multiplayer agent harness where every person and every room gets its own memory, files, keychain, crons and durable sandbox, on a harness you choose.
DeepSeek Harness: an agent runner where everything is a plugin
deepseek-ai/deepseek-harness
DeepSeek published its own agent harness, dsh, and it collected six figures of stars in two days. It is also a developer preview that promises to break.