One week after @fosdem.org, we’re still thinking about the conversations. Lots of interest in local models, privacy-first AI, and tools like llamafile and any-llm. Builders were curious, thoughtful, and honest. Thanks to everyone who stopped by 💙

Multilingual safety isn’t solved by scaling bigger models. any-guardrail now supports Alinia, a security model trained with real adversarial examples and cleaner data to detect prompt injection and policy violations across languages. See the integration: link.mozilla.ai/alinia-any-g...

A new paper with the team behind Shelf (by koodos), @johnpdickerson.bsky.social, @jad.bsky.social, @apuchitnis.bsky.social, and collaborators argues that personal context is becoming the next identity layer, and why preserving user agency matters. Paper link: link.mozilla.ai/beyond-digit...

What if “safe” AI isn’t safe in every language? @royapak.bsky.social and Daniel Nissani explore how AI safety guardrails behave outside English, and why language can affect fact-checking and risk detection in humanitarian settings. See how the evaluation was done: link.mozilla.ai/multilingual...

Agent infrastructure breaks in the gaps: authentication, rate limiting, logging, audit. mcpd Plugins formalize that in-between layer with a fixed execution order; no agent code or tool-servers changes required. Read the plugin configuration: link.mozilla.ai/mcpd-plugins...

@fosdem.org 2026 is almost here (Jan 31–Feb 1)! If you’re in Brussels, stop by the Mozilla booth and say hello. @toto.space, Nathan Brake, and @aittalam.bsky.social will be there. See you this weekend 👋

Tomorrow is the day! Nathan Brake, Senior Machine Learning Engineer at Mozilla.ai, joins GitHub’s Open Source Friday for a live conversation on building AI systems that can adapt as models and providers evolve. Tune in live: gh.io/anyllm

The State of AI report from OpenRouter and a16z captures API-based usage well. But many small models run locally, outside APIs. any-llm-platform helps surface that local activity, adding context the data can’t show. Full breakdown:

An invitation worth sitting with. AI is shaping more of our daily tools than we realize. The choices around agency openness and community will shape what comes next and who it works for.

encoderfile in action 👀 Thanks to Fahd Mirza for publishing a hands-on video that installs and tests encoderfile, showing how transformer encoders can be packaged into standalone executables. Watch here: youtu.be/QMi7Z4kqmBk?...

any-llm gives you one API across LLM providers. > Same code. > Different models. > Easy switching when you need it. The cookbook walks through setup, keys, and generating text across providers. Explore the cookbook: link.mozilla.ai/any-llm-cook...

On Jan 30, Nathan Brake, Senior Machine Learning Engineer at Mozilla.ai, will be joining GitHub’s Open Source Friday. It’s a live conversation about building AI systems that stay adaptable as models and providers evolve, and how open tools help preserve choice. Join us live here: gh.io/anyllm

Say hello to Octonous 🎊 An agent platform that runs real tasks across your apps with clear visibility and approvals. Beta opens in February. Join as an early user: link.mozilla.ai/octonous

AI-generated code has moved from meme to mainstream. Open source runs on trust, but most projects still don’t say how AI was used, if at all. We think that needs to change, and we’re testing a simple way to do it. See the approach:

any-llm’s managed platform is now in open beta and free to use. 🎉 Client-side encrypted API keys, one virtual key across providers, and usage + cost tracking for cloud and local LLMs. Sign up and test it out: link.mozilla.ai/any-llm-plat...

Running AI locally can feel complex. llamafile packages the model, server, and runtime into one executable that runs on your machine. Try the quickstart: mozilla-ai.github.io/llamafile/qu...

Closed AI systems are winning because they’re easier to use. The gap isn’t capability or values. It’s usability and integration. @mozilla.org CTO, Raffi Krikorian, lays out why fixing developer experience is key to moving from AI renters to AI owners — while the defaults are still being set.

any-llm-gateway adds production controls on top of any-llm: • Shared budgets with scheduled resets • Virtual API keys with metadata and expiration • Token and cost tracking per request • Docker-based deployment with Kubernetes-ready probes Try it out: link.mozilla.ai/any-llm-gate...

mcpd Plugins add a request and response pipeline for tool calls. Plugins run as external binaries over gRPC and execute in a defined category order. Read the full write up: link.mozilla.ai/mcpd-plugins

Teams manage MCP servers in different ways today. Where does MCP setup usually slow down? • Connecting IDEs • Managing env vars • IDE configs drift • Still figuring it out If you’re managing MCP servers for a team, take a look at mcpd-proxy. Docs: link.mozilla.ai/mcpd-proxy-d...

The real advantage isn’t picking the best LLM. It’s being able to switch later without pain. any-llm is built for that. Give it a try: link.mozilla.ai/any-llm-v1

More teams are rethinking cloud-hosted LLMs. agent.cpp provides minimal, high-performance building blocks for agents in C++, built directly around the llama.cpp ecosystem. Designed for running agents locally on your own hardware. Repo: link.mozilla.ai/agent-cpp

As 2025 wraps up, we shared a look back at the year. Open-source tools shipped, agent work moved forward, and real feedback helped shape our direction along the way. You can read the full recap here: link.mozilla.ai/2025-year-in...

LLMs don’t give clear warnings. > Edit a prompt. > Change the model. > Bump the version. Outputs change anyway. Not a UX issue. A systems one. How teams make prompts more stable: link.mozilla.ai/smarter-prom...

Single-binary encoders mean no runtime deps, no virtual envs, no network IO, and repeatable outputs. That’s the core idea behind Encoderfile. Read more here: github.com/mozilla-ai/e...

You don’t need the cloud to build useful agents. This cookbook shows how to run an agent fully offline: • local LLM (via Ollama) • local Python tools • read + write files on your machine Privacy-first by default. See the walkthrough: link.mozilla.ai/any-agent-lo...

JupyterLite now supports any-llm-gateway as a backend for its AI static front-end. any-llm-gateway routes LLM requests through a single gateway rather than tying usage to one provider. Integration details: link.mozilla.ai/jupyterlite-...

2025 at Mozilla.ai was about building in the open. New open-source tools, IRL workweeks, community events, and early platform launches all shaped how our stack and team evolved this year. Our Year in Review shares what we shipped and what’s ahead. Full blog post: link.mozilla.ai/2025-year-in...

Unlimited LLM access can get expensive. Over-restricting access slows teams. any-llm-gateway sits in between: • Usage tracking • Budget enforcement • Works across providers Read the docs: mozilla-ai.github.io/any-llm/gate...

Agent setups across local, CI, and production often rely on startup scripts, environment variables, and process control hacks. Version pinning and updates are manual. mcpd replaces this with a single config. Docs → link.mozilla.ai/mcpd-docs

If switching LLMs breaks your product, the issue isn’t the model. any-llm keeps one stable interface while providers change underneath. Read more: link.mozilla.ai/any-llm-v1

Post-mortems take time most teams wish they could get back. We’re exploring early concepts for the Agent Platform, including a Post Mortem Generator Agent that could turn incident data into clearer, faster-to-create summaries. Preview the early direction: link.mozilla.ai/post-mortem-...

More people are testing local LLMs lately. Where do you run most of your models? • Local setup • Cloud APIs • Hybrid • Learning the landscape Tried llamafile yet? Check it out here: link.mozilla.ai/llamafile-repo

Key benefits of any-guardrail 👇 • One API for many guardrails • No custom preprocessing • Easy model switching Get started with any-guardrail: link.mozilla.ai/any-guardrail

Enterprises run on the JVM, and now polyglot AI agents can too. This blueprint explores how WASM on the JVM runs Rust, Go, Python, and JavaScript agents side-by-side with strong isolation and enterprise-grade observability. Full blog post: link.mozilla.ai/polyglot-ai-...

Say hello to @toto.space, our newest software engineer at Mozilla.ai! He’s into coffee, photography, and open-source. He’s joining the team working on the any-llm managed platform to help simplify how teams use different LLM providers.

Roadmap questions can take up a big share of a PM’s day. We’re testing early concepts for the Agent Platform, including a Roadmap Agent that gives plain-English answers from roadmap details. More here: link.mozilla.ai/roadmap-agent

We’re opening the alpha for the any-llm Managed Platform! 📢 It’s a zero-knowledge vault and usage tracker for OpenAI, @anthropic.com, Google, and more. Full blog post: link.mozilla.ai/any-llm-mana...

What makes choosing an LLM the most difficult for you? Is it price, quality, or speed? If LLM routing would help you compare models more easily, the full post is here: blog.mozilla.ai/the-challeng...

The video recording of Davide Eynard’s (@aittalam.bsky.social) “Build Your Own Timeline Algorithm” talk at SFSCON is now available. He covers how to build a local timeline algorithm with Mastodon.py, llamafile, and marimo. Full session here: www.sfscon.it/talks/build-...

DataFest Tbilisi (Nov 27–29) is a wrap! Raz Besaleli, Davide Eynard (@aittalam.bsky.social), and Irina V. each led sessions on building stronger AI systems, open-source agents, and applying real software rigor to AI. Thanks to everyone who joined — see you next year!

We’re sharing an early look at concepts for the Agent Platform. The Agent Showroom includes a few prototype agents — call prep, post-mortem drafts, roadmap support — to help us learn what real teams need. See the early ideas: https://link.mozilla.ai/agent-platform

Three reasons any-llm makes development easier. • One API for cloud and local models   • Async-first design for real workloads   • Stable interface that survives provider changes   Get started on GitHub: https://link.mozilla.ai/any-llm-repo

DataFest Tbilisi begins tomorrow! Raz Besaleli, Davide Eynard (@aittalam.bsky.social), and Irina Vidal will all be speaking — covering AI resilience, software reliability, and how to run open source agents on your own terms. Details here: https://www.datafest.ge/

Openness and human agency don’t need to disappear in the age of AI. There’s a blueprint for building technology that keeps those principles at the center. Check out Mark Surman’s blog post outlining the evolving mission to do for AI what Mozilla did for the web.

Encoderfile v0.1.0 is here. A new way to deploy encoders as single-binary executables with no runtime dependencies. Faster, leaner, and built for correctness. Full details here: https://link.mozilla.ai/encoderfile-v010

The lethal trifecta is what breaks most guardrails. Tools. External data. Language input. A tough mix to secure. See the full analysis: https://link.mozilla.ai/open-source-guardrails

We built mcpd-proxy to make MCP tools easier to use in VS Code and Cursor. One shared setup. All your MCP servers. Read the full post: https://blog.mozilla.ai/mcpd-proxy-centralized-tool-access-for-ai-agents-in-vs-code-cursor-and-beyond/?utm_source=bluesky

If you work with MCP tools in VS Code or Cursor, you might appreciate what we’re sharing tomorrow. A small improvement that smooths out a familiar pain point.

On Nov 7, Davide Eynard (@aittalam.bsky.social) presented at #SFSCON on building a personal, local timeline algorithm with Mastodon.py, llamafile, and marimo. You can now view the slide deck here: https://www.sfscon.it/talks/build-your-own-timeline-algorithm/ Recording coming soon. 👀