The dev stack is being re-poured for a user that never sleeps.

A version control system written in Rust (Apache-2.0, public beta at 0.99.0) that an agent drives instead of git. `curl -fsSL oak.space/install | sh` on macOS or Linux; `cargo install oakvcs-cli` or a signed exe on Windows. The point is virtual mounts: content-addressed lazy loading means an agent can edit any repo in seconds without a full clone, so a fleet can fan out across one project without fighting worktrees or downloading everything. The unit of work is a branch-per-session, and branch descriptions replace per-commit messages. Reach for it when your bottleneck is agents waiting on clones and worktree juggling, not when one human is committing all day — git is still fine for that. The honest caveat: it's 0.99 beta, the repo says it was "written almost entirely using AI with human oversight," and you're putting your history in a format the rest of your tooling doesn't speak yet.

A self-hosted workspace from Block, Inc. (Apache-2.0, v0.3.29) where humans and AI agents share the same channels — built on a Nostr relay you run, so every message, patch, code review, CI event, and approval is a signed event in one audit log. Agents can open repos, send patches, review each other's code, run workflows, and orchestrate other agents, all in threads you can search alongside the human conversation. A feature branch becomes a room where the diff, the CI run, and the decision live together. Use it if you've got agents doing real work and your current trail is a Slack bot plus logs scattered across five tools. Tradeoffs: it's early (138 stars), the local stack is heavy (Docker, Rust 1.88+, Node 24+, pnpm), and betting your team's record-of-truth on a self-hosted Nostr relay is a bet — a deliberate one, but a bet.

Tencent's Rust sandbox (6.4k stars) for AI agents: instant, concurrent, isolated runtimes so an agent can execute the code it just wrote without putting it on your host. This is the execution layer of the agent's own stack — distinct from the defensive sandboxes we've covered (which assume the agent is hostile and box it); the framing here is enablement, giving the thing that never stops generating a fast place to run things in parallel. Reach for it when your agent generates-and-runs whole programs and hand-rolled Docker-per-task or a hosted code-exec API is your current answer. The caveat that matters: the license reads as "Other" with no SPDX identifier — read the actual terms before you build on it, and Tencent provenance will be a factor for some shops.

An Apache-2.0 Python memory layer (85 stars, v1.0.0) for Claude Code, Cursor, and Codex that keeps decisions, lessons, and project facts in a single local SQLite file — no server, no cloud, no API keys. Retrieval is hybrid: BM25 plus multilingual vector embeddings plus an entity graph, fused with reciprocal-rank-fusion and an optional cross-encoder rerank, at ~35ms warm recall. It plugs into the agent's lifecycle over MCP, injecting relevant memory before the agent responds. Use it when your agent keeps re-litigating decisions it already made three sessions ago and you don't want a cloud memory SaaS holding your context. The pitch is "find the right memory, not the closest one" — worth confirming that holds on your own repo, and it's a small project, so treat v1.0.0's stability contract as a promise, not a track record.

A free, open-source local gateway (from SouthForge) that sits in front of your MCP servers. Instead of loading every tool from every server into context on each request, it exposes three meta-tools the agent searches on demand — the full catalog is still there, it just stops sitting in the prompt every turn. The vendor's benchmark claims ~90% fewer total tokens and 97% less tool overhead at the same task success rate. Keys go in the OS keychain, not a config file; runs on Windows, macOS, and Linux. Reach for it once you've wired up eight or ten MCP servers and noticed a real share of your context is just tool definitions. It doesn't replace anything — it brokers what you already run. Tradeoffs: it's another hop in the loop, search-on-demand adds a round trip before a tool call, and the token numbers are the author's own.