Did you onboard your AI agents, or just hire them?

Questions

Which AI agents does Vivari work with?

The harnesses you already use: Claude Code, Codex CLI, and Cursor CLI today. Through those same harnesses, Vivari also runs local models and custom OpenAI-compatible endpoints, so your fleet is never locked to one vendor. More runtimes are planned, including Hermes by Nous Research. The workspace is engineering-first, but it manages any work you hire agents for: code, research, operations, content.

Is Vivari only for coding agents?

No. Shipping code is the first job most teams hire agents for, and Vivari Guard is purpose-built for change safety in code. The workspace itself manages fleets doing any knowledge work.

Is Vivari overkill if I only run one or two agents?

You do not need a hundred agents to have a management problem; you have one the moment a single agent gets meaningful access to your systems. One agent with write access already needs context on arrival, scoped access, review on risky changes, and a record of what it did. That is what the Solo tier is for, and the fleet features will be waiting when you grow.

Does my code leave my machines?

Vivari never uploads or mirrors your repositories to us, and we never train on your code. Vivari Guard's review runs entirely on your own machine, with no network call for the analysis. The security page maps the full picture, including what your chosen model provider sees.

Is Vivari reactive or autonomous?

Autonomous, accountably. Most tools wait for a prompt. The Conductor runs on a heartbeat: it watches the fleet, drafts and curates between your prompts (one model family proposes, a different one reviews), and brings you finished work. All under the same permissions, review, and audit trail as everything else. In auto mode a deterministic rule engine and risk evaluator decide what runs and what escalates to you: no model clears its own risky actions.

Won't better models make all this unnecessary?

The assumption is that a smarter model shrinks the management problem, and in practice it grows it. A weak model writes clumsy work you catch on sight; a powerful one writes convincing work, risky in ways you miss on first read, and faster than anyone can check. Checking is the half that does not speed up with the model. So every model upgrade raises the value of the layer around it, and Vivari is built so you can swap the model in the day a better one ships.

Couldn't we just build this ourselves?

Yes, and the parts are good: memory stores, sandboxes, observability, policy engines. A stack of parts still cannot answer the questions that matter, because each tool holds its own fragment of the story. The management model only exists when identity, context, permissions, evidence, and history are one system; start with the questions your stack cannot answer today.

When does Vivari launch?

Curated early-access cohorts open fall 2026. Cohorts stay small because every workspace gets guided setup and calibration against your own repositories.

Why should I believe any of this?

You don't have to take our word: the pre-merge log above is pulled live from this site's own repository, and every entry carries the commit and evidence behind its verdict. At launch, the receipts page goes live: the Vivari fleet sessions behind this site's build, and every deploy risk-scored by Vivari Guard. The manifesto explains the thesis. Then flip the question around: play The First Week: eight decisions, ninety seconds, a grade for how you treat your own agents.