--- title: "What is Cerevisor?" description: "A two-minute orientation to Cerevisor, what it does, who it's for, and the three ideas you need to know before you build anything." slug: getting-started/what-is-cerevisor section: getting-started canonical_url: https://cerevisor.com/docs/getting-started/what-is-cerevisor last_verified: 2026-05-18 last_verified_version: "1.2.0" updated_at: 2026-05-18T15:08:18.053416+00:00 --- Cerevisor is a desktop application for designing and running **multi-agent AI workflows**. You drag agent cards onto a canvas. You give each agent a role (researcher, developer, reviewer, writer), a small set of skills (focused capabilities), and a model (Claude, GPT, a local Ollama model, Cursor, whatever you have). You connect them in the order they should work. Then you press Run, and they actually do the work, calling tools, reading files, writing files, asking each other for what they need. That's the whole product. ## Who it's for Knowledge workers. People whose work involves chains of reasoning, research, drafting, reviewing, and shipping, and who have started to suspect that a single chat with a single model isn't the right shape for what they're doing. You don't need to be a developer. You don't need to write code. You do need to be willing to think carefully about how *you* would break a job down, because the workflow you design is the plan the agents execute. ## The three ideas If you understand three things, the rest of the app is just buttons. ### 1. A **workflow** is a left-to-right sequence of waves. Each column on the canvas is a wave. Agents in the same wave run in parallel. Agents in later waves wait for earlier waves to finish, then read what those agents produced. This means: don't think about agents one at a time. Think about *who needs what from whom*. The columns enforce the order. ### 2. An **agent** is a role + skills + a model + instructions. The role (e.g. *researcher*) sets the agent's defaults. Skills (e.g. *recency-research*, *docx*) are focused capabilities the agent can call on. The model (e.g. Claude Sonnet, local Llama, Codex) is the engine. The instructions are how *you* tell the agent what you want from it in this specific workflow. Everything else on the agent card, output definitions, input/output files, tool permissions, is refinement. Start with role + instructions and add the rest when the agent needs them. ### 3. **Skills** are reusable, focused capabilities. A skill is a small package that teaches the harness one specific thing well, how to format a Word document, how to do recency-aware web research, how to publish a blog post, how to build a Remotion video. Cerevisor scans `~/.claude/skills/` and `~/.cerevisor/skills/` on startup and merges everything it finds. You assign skills to an agent by dragging them onto the agent card. You can also create your own in the **Skill Workshop**. If you've used Claude Code, skills will feel familiar; they're the same files, in the same folders. ## What Cerevisor is **not** - **Not a model.** Cerevisor doesn't ship its own LLM. It connects to whatever provider you tell it to use. - **Not tied to one provider.** Anthropic, Gemini, Claude Agent SDK, OpenAI-compatible endpoints (Ollama, OpenRouter, vLLM, LM Studio, and friends), OpenAI's Codex CLI, and Cursor are all first-class. You can mix providers within a single workflow. - **Not a cloud service.** Everything runs locally on your machine. Memory, profile, run logs, and audit trails all live under `~/.cerevisor/`. No telemetry. No data leaves your device unless *you* configured a provider that requires a network call. - **Not Claude Code.** Cerevisor can *export* a workflow as a prompt that you paste into Claude Code if you prefer the CLI experience: but the default Run button executes natively against the provider you've chosen. ## What it gives you that a chat doesn't - **Parallelism.** Three researchers can hit the web at the same time while two writers wait for them. - **Specialization.** A reviewer agent on a small fast model can vet the output of a senior agent on a more expensive model: without you switching tabs. - **Repeatability.** A saved `.cerevisor` file is the workflow. Run it once, run it a hundred times, schedule it for every Monday at 9am. - **Memory.** The app remembers what you tend to ask for, what worked, what didn't, and applies that context to every run: without you re-explaining your preferences each time. - **Provenance.** Every run writes an audit trail you can read later. You can see which agent called which tool, when, with what input, and what it produced. ## Where to go next → [Install Cerevisor](./install.md)