The agent is being onboarded like an employee.

A single self-contained binary (Apache-2.0, .NET runtime embedded) that reads, edits, and renders Word, Excel, and PowerPoint files headlessly — no Microsoft Office anywhere on the machine. The verb set is built for a machine operator, not a human: view a document as outline, text, HTML, or screenshot so the agent can see what it's editing; get/set/add/remove by path-based element addressing; merge for template filling; validate before you ship. Excel gets 350+ functions with formula auto-evaluation; a built-in MCP server registers itself with Claude, Cursor, VS Code, or LM Studio in one command. Install is brew, npm, or one curl.

Reach for it the day your agent's work product is a document instead of code — the weekly report, the invoice run, the deck nobody wants to make. Delete the python-docx + openpyxl + python-pptx trio and the glue code that never handled tracked changes anyway. Tradeoffs: this is an established rocket, not a launch — created in March, 9,250 stars, trending again today — and the README's "world's first and best Office suite for AI agents" tone oversells what is, concretely, a very good headless document CLI. It's also a monolithic C# binary: you adopt the whole suite, not a library you compose.

An open-source security layer for agents, named for the banking control it implements: the one who makes isn't the one who checks. Three composable pieces. mc scan statically flags what an agent's tools could do — data deletion, financial transfers, shell access, secret exfiltration — and sorts them into risk tiers. An embedded library wraps those tools in deny-by-default, role-based access control at the call boundary. An optional self-hosted server adds human-approval workflows and an audit log that's hash-chained and Ed25519-signed — alter a row and verification breaks, and bundles export for offline verification without trusting the machine that produced them. Connectors ship for LangChain and the Claude Agent SDK, with TypeScript and Python SDKs.

This is the enforcement-beats-instruction pattern again — the same layer deptrust and agent-vault work at — applied to segregation of duties: the agent acts only through a role, runs only what the role grants, and cannot sign off on itself. Reach for it when agents touch consequences: money, production, customer data. Delete the system-prompt paragraph asking the agent nicely to be careful. Tradeoffs: created June 11, 41 stars — a fresh bet, pushed daily but young; and the license is split — scanner, SDKs, and connectors are Apache-2.0, the server is AGPL-3.0 with commercial licensing on offer. Read the boundary before you build on it.

Mozilla AI's open-source, OpenAI-compatible LLM gateway you run yourself (Apache-2.0, one Docker command to start). One endpoint fronts 40+ providers; the control surface is the point: provider keys stay private on the server, clients get scoped, revocable gw- keys, and per-user and per-key budgets are enforced before a request runs — a hard stop, not a surprise on next month's invoice. Usage is queryable at /v1/usage, and there are optional request-level guardrails (prompt-injection checks) plus sandboxed code execution and web search if you want the gateway to carry the tools too.

Reach for it when several agents, teammates, or apps share model access and you want the bill and the blast radius scoped per client — the natural companion to yesterday's agent-vault, one layer up: that hid the credential from the agent, this meters what the credential is allowed to spend. Delete the raw Anthropic key pasted into every agent's env, or the tolerated LiteLLM config if budgets were the feature you actually wanted from it. Tradeoffs: 149 stars, created in April, in a category where LiteLLM has years of head start and a much longer feature list; and the hybrid mode connects to the hosted otari.ai platform — standalone mode is the self-owned path, choose it deliberately.

An MIT per-node telemetry agent from Last9 that ties GPU hardware metrics to the workload responsible: on Kubernetes it attaches pod, namespace, and deployment; on Slurm it appends job ID, username, account, and partition. Vendor coverage is genuinely broad — NVIDIA via NVML/DCGM (T4 through GB200), AMD MI300X/MI325X via amdsmi, Intel Gaudi via hl-smi — and everything leaves as standard OTLP metrics and logs to any backend you already run. The README says it outright: no Last9 lock-in in the agent, no account required. Install is pip install l9gpu on bare metal or a Helm-deployed DaemonSet on K8s.

Reach for it when the cluster is shared, the monthly GPU line item is large, and "which team is burning the H100s" is answered by asking around. Delete the nvidia-smi-in-a-tmux-pane ritual and the spreadsheet mapping node names to teams. Tradeoffs: 53 stars, created in April — young, though the org behind it is an observability company with skin in the game; and one gauge per GPU per metric per 60-second cycle is chargeback resolution, not profiler resolution — this tells you who, not why slow.

A local visualization tool (Apache-2.0 by its README) that renders a language model's internal activations in real time: a Jacobian lens reads nine layers and displays the internal state as words, sized and faded by activation strength, amber while the model reads your input, blue while it generates — so you watch it settle on an answer before the first token appears. It ships pre-fitted lens weights for Qwen3.5-4B, runs as a Python server plus browser UI, and there's a hosted replay of example sessions if you want the idea without the ~9GB download. Any HuggingFace decoder works if you fit your own lens; a 27B variant exists for bigger cards.

Anthropic's global-workspace paper was on today's front page at 383 points; this is the same instinct at hobbyist scale, running on your GPU tonight. Reach for it to build intuition about what a small local model actually does with your prompt — where it commits, what it considers and discards. It replaces reading logprobs and guessing. Tradeoffs: created yesterday, 44 stars — a fresh bet; needs ~10GB VRAM or 16GB Apple Silicon; the GitHub API reports no license even though the README declares Apache-2.0 (a LICENSE file would fix that); and a lens is a learned approximation of the model's state, not a readout — treat it as intuition, not evidence.

Off the thread but worth knowing: Pulpie (usefeyn.com) is a family of open-weights encoder models that strip boilerplate from raw HTML and return clean markdown — pip install pulpie, weights on Hugging Face, and the 210M "Orange Small" matches the leading extractor's quality at roughly 20x less cost; if agents are the new employees, this is the mailroom. Ternlight (github.com/soycaporal/ternlight, MIT) is a 7MB ternary-weight embedding model on npm that embeds in ~5ms in the browser, CPU-only, zero network — semantic search with no server at all. And Riddle (github.com/MaximeRivest/Riddle, MIT, 881 stars in two days) turns a reMarkable Paper Pro into Tom Riddle's diary — write with the pen, the page drinks your ink and answers in a flowing hand. Not a product category, just the weekend's best reminder that the ambient-AI hardware was already in your bag.