Frequently Asked Questions

Honest answers about AI agents and Cerevisor: what it does, which models it works with, where your data goes, and what it costs.

AI Agents and Harnesses, Explained

What is an AI agent?
An AI agent is a language model given a goal, a set of tools, and the latitude to decide its own steps. Instead of answering one question and stopping, it can plan, call tools (search the web, read a file, hit an API), check its own work, and loop until the task is done. A chatbot responds; an agent acts. In Cerevisor you compose agents on a visual canvas and decide exactly how much autonomy each one gets.
What is an AI harness?
A harness is the layer between you and the raw AI models that makes a fleet of agents actually runnable. It orchestrates who does what, passes context between steps, calls tools and MCP servers, keeps humans in the loop, and records everything that happened. Without one you are gluing prompts together by hand and hoping it holds. Cerevisor does this on your desktop, across every AI tool you use, so orchestration, the run record, and provider switching come out of the box.
What is the difference between a chatbot and an AI agent?
A chatbot is one model and one prompt: you ask, it answers, the conversation is the product. An agent is given a goal and the tools to pursue it across many steps, often handing off to other agents along the way. Chatbots are great for quick answers; agents are for work that has a process: research, drafting, review, routing, follow-up. The moment your task needs more than one prompt, you are really building a workflow, which is what a harness is for.
What is an AI skill?
A skill is a reusable, packaged capability an agent can call: a saved prompt, procedure, or mini-workflow that encodes how to do one job well, such as writing a brief, extracting data from an invoice, or reviewing code against a checklist. Instead of re-explaining the task every time, you build the skill once and any agent can use it. Cerevisor includes a skills library so your team's best procedures become assets you reuse.
What is a multi-agent workflow?
It is several specialized agents working together on one job, each handling the part it is best at. A planner breaks down the work, an extractor reads source material, a writer drafts, a reviewer checks, and a router decides what runs next. This mirrors how a team operates and produces better results than asking one model to do everything in a single prompt. Cerevisor's canvas is built for exactly this: you wire the agents together and watch the work flow between them.
What is MCP (Model Context Protocol)?
MCP is an open standard for connecting AI models to tools and data: filesystems, browsers, databases, and your own internal systems. A model that speaks MCP can call those tools without anyone writing custom glue code for each one. Cerevisor's runtime is MCP-aware out of the box, so you plug in an MCP server and your agents can use it as a tool, with every call captured in the run's audit trail.
What does 'human in the loop' mean?
It means a person stays in control at the points that matter: approving an action before it runs, reviewing output before it ships, or being asked to decide when the agent is unsure. Full autonomy is rarely what you want for real work; you want speed where it is safe and a checkpoint where judgment is required. Cerevisor lets you place approval and review steps anywhere on the canvas, so agents move fast on the routine parts and pause for you on the consequential ones.
What is a local-first AI app, and why does it matter?
Local-first means the software runs on your own machine and your data stays there rather than passing through a vendor's servers. For AI work that matters because your prompts, your provider API keys, and your run history are sensitive. Cerevisor runs locally and talks straight to whichever AI providers you choose. Your keys stay encrypted on your device, your prompts and outputs go only to the providers you chose, and the app keeps its usage data on your device.

Building and Automating with AI

How do I build an AI agent?
At the simplest level you give a model a clear goal, the tools it needs, and instructions for how to behave, then test it on real tasks and refine. The hard part is rarely the first prompt; it is the orchestration around it: handing context between steps, calling tools reliably, knowing when to stop, and seeing what went wrong. A harness gives you that scaffolding so you focus on the logic. In Cerevisor you drag an agent onto the canvas, pick its model, give it instructions and tools, and run it, with no glue code to write.
Do I need to know how to code to build AI agents?
Cerevisor's canvas is built for people who think in workflows: you compose agents, connect them, and add approval steps visually. If you do write code, nothing gets in your way, and you can drop in custom logic and connect your own tools via MCP. For most people the real constraint is iteration speed on the workflow, and that is what the canvas speeds up.
How do I automate tasks in my work and life with AI?
Start by picking one repeatable task that eats your time and has a clear input and output: triaging an inbox, turning notes into a summary, drafting first-pass replies, pulling data out of documents. Map the steps you take by hand, then assign each step to an agent and put a review checkpoint wherever a mistake would be costly. Cerevisor is built for exactly this loop: build the workflow once, run it on demand, inspect what happened, and improve it, so a one-off automation becomes something you can trust to repeat.
How is this different from automation tools like Zapier or Make?
Classic automation tools connect apps with fixed if-this-then-that rules, and they are excellent when the steps stay the same every time. AI agents handle the messy parts: reading unstructured text, making judgment calls, writing prose, deciding what to do next. Cerevisor is for the work that needs reasoning at each step, and it keeps a human in the loop where the decision actually matters.
Where do I start with my first agent?
Download Cerevisor and choose 'Guide me through my first project'; you will have a multi-agent workflow running within minutes before you build anything of your own. From there, pick one real task, copy a template close to it, and adjust. You learn the concepts (agents, skills, tools, review steps) by running them.
What kinds of work can AI agents handle today, and what should stay manual?
Agents are strong at research, drafting, summarizing, extracting and reformatting data, first-pass review, and routing work to the right next step. They are still unreliable on anything needing real-world judgment, high-stakes irreversible decisions, or guaranteed accuracy without a human check. The practical answer is to let agents do the volume and keep yourself on the checkpoints, which is the entire reason Cerevisor makes human-in-the-loop review a first-class part of every workflow.

About Cerevisor

What is Cerevisor?
Cerevisor is a desktop app that runs a team of AI agents for you. Design the team on a visual canvas or describe it in chat, give each agent its own model, check the plan before it runs, and keep a record of every run on your machine. It works with the AI tools you already use, from Claude Code and Codex to API models and local Ollama.
Who is it built for?
Business workers and power users who already run several AI tools for their work: transformation leads, analysts, client-facing teams, founders, and anyone who spends their day carrying context between tools and checking the results. It suits people who think in workflows: a weekly client update, a research brief, a board deck. It works without code, and it welcomes your code when you have some.
Is this a SaaS product or a desktop app?
Desktop app. Cerevisor runs on Windows and macOS, on your machine. You install it and start: your runs go from your machine straight to your providers, with your own keys, which stay with you.
How do I get started?
Download the installer for Windows or macOS, choose 'Guide me through my first project', and connect a model: sign in with a plan you already have, paste an API key, or use a local Ollama model. You will be running your first multi-agent workflow within minutes. The app is free to start, with a 7-day trial of everything.

Privacy and Security

Where does my data go when I run a workflow?
From Cerevisor on your machine straight to the AI provider you chose for that agent: Anthropic, Gemini, an OpenAI-compatible endpoint, Ollama, Claude Code, Codex, or whatever else you wired in. The path is direct, so your prompts, runs, and outputs stay between you and your provider. The app keeps its usage data on your machine and sends none of it to Cerevisor.
Where are my provider API keys stored?
Locally on your machine, encrypted with your operating system's secure storage (Windows DPAPI or the macOS Keychain). They stay on your device, and only you can use them.
Are my workflows or runs uploaded anywhere?
No. Workflows live as .cerevisor files and worlds live as .cerevisor-world files on your disk. Run artifacts (NDJSON audit logs, outputs, intermediate state) stay local. Nothing is synced to Cerevisor unless you explicitly export and share it yourself.
What about audit trails for compliance and reviews?
Every run writes an append-only record on disk: 38 kinds of fact, including permission decisions, approvals, tool calls, files touched, web requests, and cost. Audit logs are NDJSON files in ~/.cerevisor/audit-logs, so you can answer 'why did the agent do that?' from data instead of memory, and pipe the files into whatever compliance or analytics tooling you already use.
Can I run it offline with local models?
Yes. Point an agent at Ollama or any OpenAI-compatible local endpoint and the workflow runs without an internet connection. You can also mix local and cloud providers in the same workflow when you want privacy on some steps and frontier capability on others.

Models, Providers, and MCP

Which AI providers does Cerevisor support?
Eight provider types. Cerevisor runs the agent itself on Anthropic, Gemini, and any OpenAI-compatible endpoint (OpenAI, OpenRouter, Groq, Together, Azure OpenAI, LM Studio, vLLM, and local Ollama). It also drives agent tools you may already pay for: Claude Code (bundled), Codex, Cursor, Antigravity, and Grok Build, signed in with your own plan or key. Adding a new OpenAI-compatible endpoint is a quick configuration change.
Can I mix providers inside one workflow?
Yes, and that is one of the main reasons to use Cerevisor. Use a frontier model for planning, a fast cheap model for extraction, a local model for sensitive steps, and a specialist model for code review, all in one workflow. Each agent picks its own provider.
How does MCP work in Cerevisor?
On the Pro and Founding member plans, plug in any MCP server (filesystem, browser, database, custom internal tools), or pick one from the built-in catalog, and your agents can call it as a tool without you writing glue code. Every MCP call shows up in the run record like the rest of the run.
Will Cerevisor be outdated in six months as models change?
The model landscape changes fast and Cerevisor is built for that. It keeps you free to switch providers, and adopting a new endpoint or model is a quick config change. Cerevisor is updated regularly to keep up with new providers, agent tools, and MCP servers, so your workflows keep running on the best model for the job.

Pricing and Value

What does it cost?
Free to start, with a 7-day trial of everything, so you can try every paid feature on real work. The Free plan is for personal use, students, teachers, and nonprofits; work for a business, an employer, or paying clients needs a paid plan. There are two paid plans, Pro and Founding member; Pro leaves out the two experimental features, Operator and the team conversation. See the pricing page for current prices. Both are built around individuals and small teams, with simple self-serve checkout. You pay for Cerevisor; you pay your AI providers separately for model usage, or use a plan you already have.
Why pay for Cerevisor instead of using ChatGPT or Claude directly?
One-prompt chat tools are great when one model with one prompt is enough. Real workflows usually need more: a planner, an extractor, a writer, a reviewer, and a router; tools and MCP servers; and a record of what happened. Bolting that onto a chat UI means reinventing the orchestration layer. Cerevisor ships that out of the box, and it can still use your ChatGPT or Claude plan to do the work.
Why use Cerevisor instead of stitching together our own multi-agent setup?
You can. Going DIY means writing glue code, debugging context plumbing, building your own audit trail, wiring up MCP servers, and maintaining all of it. Teams spend months on this and end up with something that works for one workflow and is fragile beyond it. Cerevisor gives you a working canvas, multi-model orchestration, a local run record, and MCP tools on day one. You bring your own provider keys and your own logic; the plumbing stays predictable.
Is there a price lock?
Yes. Join as a Founding member and your price stays the same for as long as your subscription is active, even if standard prices change later. Anyone who subscribed before Cerevisor 3.3.0 is already a Founding member. Cerevisor is in active development, and feedback from members shapes the roadmap.

Why Cerevisor

Why use Cerevisor instead of hiring AI engineers?
Hiring is the right move when you have a long-running product surface that demands deep custom infrastructure. For most teams the constraint is iteration speed on workflows. A senior AI engineer in Western Europe runs EUR 80,000-150,000/year before benefits and equity, and 6-12 months to ramp. Cerevisor lets a small team run multi-agent workflows on day one, with the option to layer in custom code where it actually matters. It works best alongside your engineers.
When do I need more than a single AI chatbot or a one-agent tool?
One-agent tools are fine when one model with one prompt is enough. Real workflows usually need more: an architect to plan, an extractor to read, a writer to draft, a reviewer to check, and a router to decide what runs next. Bolt those onto a chat UI and you end up reinventing the orchestration layer. Cerevisor is built for that from the start: visual canvas, a model per agent, skills library, and a record of every run.
What makes Cerevisor different from other agent platforms?
Three things. First, it runs locally on your machine and keeps its usage data there, so prompts, runs, MCP traffic, and provider keys stay with you. Second, it is provider-neutral by design: eight provider types, from Claude Code and Codex to Gemini and local Ollama, mixed per agent in one workflow. Third, every run writes a record on your machine, so you can answer 'why did the agent do that?' from data. And because Cerevisor sells no models, it treats every vendor equally.

Support, Updates, and Lock-in

What happens after I subscribe? Am I locked in?
No. Cancel anytime from the Lemon Squeezy customer portal and you keep your plan through the end of the period you have already paid for. The first paid purchase has a 14-day money-back request window. Workflows and worlds live as plain files (.cerevisor and .cerevisor-world) on your machine, so even if you stop subscribing, what you built is still yours to version, share, and re-run.
Do I get updates and bug fixes after I install?
Yes. Cerevisor ships regular releases, and updates reach every install, free or paid. On Windows the app updates itself. On Mac, download the new version from the site until the Mac build is signed.
Can I share workflows with my team?
Yes. Workflows and worlds are plain files on disk, so sharing is whatever your team already does for files: a shared drive, a git repo, an attachment in chat. Provider keys stay on each person's machine, so teammates use their own credentials when they re-run what you sent.
How do I report a bug or request a feature?
Email support@cerevisor.com or use the contact form in the footer. Paid members get a direct path to the team, and feedback actively shapes the roadmap. If something breaks, we want to hear about it.