Tutti exists because every agent framework I tried in early 2026 had a different deal-breaker — security as an afterthought, no proper plugin model for tools, weak eval primitives, or a runtime too coupled to one language. So I started writing the one I wished existed. It’s open source, opinionated, and built in public.
The runtime is Apache 2.0 forever. There's no plan to lobotomise the OSS to push a paid tier. If anything Tutti-related goes away tomorrow, your score files keep running, your voices keep installing, your team keeps shipping.
Most agent systems are built by writing more Python. Tutti is built by writing less. A typed score file is reviewable, diffable, computable, and shippable as a package. Code is the last resort, not the default.
Tutti looks more like Terraform than like a notebook. Score files don't run inside-out. Voices don't monkeypatch globals. The runtime never edits its own prompts. Boring shapes are what survive Tuesday morning incidents.
Observability, evals, HITL, secret scoping, prompt-injection defences — these were on day one, not in the roadmap. If a framework treats production as the next quarter, you become the test case.
Tutti started in March 2026. Every commit since has been on GitHub. No stealth period, no rebrand, no rewrite — what you see is what was built.
I was trying to build a real multi-agent workflow and kept evaluating frameworks that almost worked. LangGraph was powerful but treated security as an afterthought. CrewAI moved fast but couldn't model a proper voice plugin pattern. Mastra was clean but missing the eval and HITL primitives I needed. None of them got the defaults right.
Sketched a typed score-file format on paper, opened a repo, and started writing. The first three voices — filesystem, github, web — landed within the first ten days. Public on GitHub from day one.
Built out the runtime, the CLI, more voices, the MCP bridge, the permission model, the prompt-injection guard. Wrote tests as I went — there's no separate 'add tests later' phase, the tests are how I know the score-file shape is right.
I'm one person. I'm not pretending to be a team. The docs say what works and what doesn't. The roadmap is on GitHub. If you try Tutti and something breaks, open an issue or DM me on Discord.
I’m an engineer based in the Netherlands. Tutti is what I’m working on full-time. I’m not a marketing team or a content person — if the docs are wrong, that’s on me, and the fastest way to fix it is to tell me.
The roadmap is public. The bugs are public. The decisions are written down so anyone reading along can argue with them. If you’ve built agents in production and have something to say about how Tutti should work, I want to hear it.
Right now, just me — Chihab. I'm based in the Netherlands. The OSS is open to contributors and a couple of people have already shipped patches, but there's no team yet. If that changes, this page will say so.
Because the existing ones each had a deal-breaker for what I needed: security as an afterthought, no proper plugin model for tools, missing eval primitives, or a runtime that was too coupled to one language. Tutti is the framework I wished existed when I started.
Discord for community questions, GitHub Issues for bugs, hello@tutti-ai.com for everything else. Security issues should go to security@tutti-ai.com — please don't open public issues for those.
Bootstrapped. No investors, no runway pressure, no incentive to lock anyone in. If that changes I'll write about it openly — the OSS commitment isn't on the table either way.
No. The whole point is that the runtime stays Apache 2.0 and stays focused on multi-agent orchestration done well. If a hosted version eventually exists to keep the lights on, the OSS will still do everything it does today.
One install. Your first agent running in 60 seconds. No signup. No telemetry.