Build the agent. Swap the model. Keep the work.
The harness is what runs under every playbook, app and pipeline in Corvic. You choose the model, connect tools over MCP, and ground the agent in the data in your room — without writing a line of code.
What the harness gives an agent
The parts a team ends up building by hand — model choice, tools, grounding, permissions — already assembled and swappable.
Configure, don't code
Model, prompt, tools and knowledge on one screen. No framework to learn.
Every model, no lock-in
OpenAI, Anthropic, Google, xAI, open source — or your own endpoint.
Compare, then promote
Run one task across models, weigh quality against latency and cost.
Tools over MCP
Web search, warehouse queries, files — called with your credentials.
Grounded in your room
Answers come from the tables and documents you attached, with citations.
Scoped by permission
A room bounds what an agent can read. Access follows the person asking.
How an agent gets built
- 1
Open a room
Attach the tables, files and connectors this agent is allowed to see.
- 2
Pick a model and tools
Start on a default, add MCP servers, then swap models whenever you like.
- 3
Put the agent to work
Chat with it, save the run as a playbook, or let it build an app.
Try it yourself
A working replica of a Corvic room, on invented data. Switch rooms, change the model, open the tabs — then take your own question to the real thing.
Sketch out your next data app.
Describe what you want to build
Then the work leaves the chat
An answer in a thread helps once. The harness hands its output to the three surfaces that keep it running.