Two ideas are converging in today's most-starred projects: making AI small enough to live entirely on constrained hardware, and building the collaboration layer that lets AI agents work as persistent teammates rather than stateless tools. Whether the trend is a 14MB model binary or a shared-memory workspace, the underlying question is the same — how do you make AI genuinely useful in the places and workflows where people actually are?
cactus-compute/needle: A Foundation Model That Fits in 14MB
Needle 2 is a 45-million-parameter model compressed into a single 14MB binary that runs in roughly 28MB of RAM. It targets edge environments — mobile apps, wearables, robotics, smart home systems — where shipping a dependency on a cloud API isn't an option. The technical approach is notable: a "Simple Attention Network" architecture with Hadamard MLPs and CQ2-bit quantization bakes everything, including weights, into one self-contained file. Tool calls return structured JSON via a byte-level grammar compiled from your schemas, and a sliding 256-token context window keeps memory usage flat regardless of conversation length. LoRA fine-tuning and a .cact export format round out the developer story.
github.com/cactus-compute/needle
github/spec-kit: Write the Spec First, Let Agents Build the Rest
GitHub's Spec Kit formalizes a workflow called Specification-Driven Development — write a detailed spec first, then hand it to AI coding agents as the authoritative source of truth for the implementation. The toolkit ships a set of slash commands (/speckit.specify, /speckit.plan, /speckit.tasks, /speckit.implement) that walk a project from initial requirements through planning and task breakdown to code generation. It's compatible with 30+ agents and is deliberately stack-agnostic, treating SDD as a process rather than a framework. The most interesting design decision is separating "managed tooling files" (kept in sync by the kit itself) from "specification artifacts" (updated only when intended behavior actually changes) — a distinction that matters a lot once a codebase drifts from its original plan.
holaboss-ai/holaOS: One Workspace for Multiple Agents
holaOS is an open-source desktop workspace that lets Claude Code, Codex, and its own holaOS agent share a single environment with a persistent, locally-stored memory. Rather than starting fresh every session, agents can read and write context that accumulates across runs — stored as editable plain files — so a project's history isn't siloed inside one tool's conversation window. The workspace adds 100+ integrations (GitHub, Slack, Linear, Gmail), a real browser panel embedded next to the agent, and bring-your-own-key support for Anthropic and OpenAI endpoints. It's built in TypeScript on Electron and ships on macOS, Windows, and Linux under a modified Apache 2.0 license.
macro-inc/macro: An All-in-One Workspace Written in Rust
Macro is an open-source (AGPLv3) team workspace that replaces Slack, Notion, Linear, and HubSpot with a single application: email, threads, docs, tasks, a canvas board, calls with automatic transcription, and a lightweight CRM, all bidirectionally linked. What makes it technically interesting is the stack — 42+ backend services and 167 libraries in Rust using a hexagonal architecture, with CRDT-based collaborative document editing synced via Cloudflare Durable Objects. The "Agents" module maintains a nightly-refreshed team memory and exposes it over MCP, so any agent connecting to the workspace has access to the company's full context. SOC 2 Type II and ISO 27001 certified for teams that need compliance out of the box.