Build with AI

Visivo includes an Agent tab and an MCP server. Use Visivo's built-in agent or connect an agent you already use, such as Claude Code, Cursor, or Gemini CLI. Both work with the same project tools.

Choose your path

Path	Where the agent runs	Who supplies inference	Start here
Your agent, your machine	Your external agent connects to local visivo serve through MCP	Your agent's provider account or subscription	Connect with MCP
Our agent, your machine	The Agent tab in local visivo serve	Your own provider key	Use your own key
Our agent, our inference, your machine	The Agent tab in local visivo serve	Visivo, paid from your account's credit	Authorize local Visivo inference
Our agent, our infrastructure	The Agent tab in Visivo Cloud	Visivo, paid from your account's credit	Use the agent in Cloud

Choose MCP to stay in your existing agent, or the Agent tab to ask for changes beside Workspace, Runs, and Dashboards. Choose your own key for provider-funded local inference, or authorize Visivo inference to use your account's credit. Choose Cloud when the project and its agent should run on Visivo's infrastructure.

Visivo inference spends credit

Both Visivo-funded paths spend account credit when the agent calls the model. Cloud builds also draw on that balance. Read AI credit and costs and check Billing before sending your first prompt. Your own provider key also costs money, billed by that provider.

You review and commit

The built-in agent and Visivo's MCP tools cannot write your project YAML files or commit changes. Every project write lands in the draft tier. Review drafts in the viewer, inspect results in Runs and Dashboards, then choose Commit yourself or discard the changes.

The tools do not expose shell execution, filesystem writes, commit, deploy, or delete operations. Draft edits can still trigger builds under your run settings; this boundary is not a promise of no queries or no charges.

An external agent may have its own filesystem and shell tools. Visivo's boundary applies to its MCP tools, not to those independent capabilities. Ask your external agent to use Visivo tools for project edits and retain its own approval controls.

Credentials and project guidance

For the local built-in agent, a provider environment key such as ANTHROPIC_API_KEY wins over agent.api_key in ~/.visivo/profile.yml. Both win over the device token saved by visivo authorize. A BYO key therefore continues to be used even when you are signed in to Visivo. See credential precedence before switching paths.

visivo init creates a project AGENTS.md for naming conventions, preferred sources, and metric definitions. Visivo also ships reusable skills. The built-in loop loads the brief and uses those skills; external agents can read the same skills over MCP. See project guidance and skills.

Shared projects and prompt injection

Object names, SQL comments, Markdown, and column names can reach the agent's context. In a shared project, another contributor can control that text. Text that tells an agent to ignore your request, reveal credentials, or publish a change is a prompt injection attempt, not a project instruction.

The built-in agent is instructed to treat project content as data. The draft-only tool boundary limits unwanted edits to changes you can inspect and discard. It does not make model output trustworthy or prevent context already sent to a provider from reaching that provider. Review project guidance and drafts, and provide only context you intend to share with the chosen inference service.

Start with a specific request: “Read the existing revenue Model and Insights. Draft a monthly revenue view using the project's definition of revenue, and leave the changes for me to review.”