Give your AI agent an automation engine.
Connect Claude Code, Cursor, VS Code, Claude on the web, or the kovca CLI. Sign in, choose a workspace and access, and let your agent work within your permissions. Every workflow edit shows up in the Studio.
Add the URL, sign in
Kovca is a remote MCP server over Streamable HTTP. Add it without an Authorization header, authenticate in your client, then choose a workspace and access in your browser. Author is the default ceiling, within what the client requested.
Claude on the web: add a custom connector with the remote MCP URL, then Connect. These instructions are not a verified compatibility matrix: OAuth has not yet been tested end to end in these clients. Client version, plan, and organization policy may affect availability. Sign-in and connection management
Your consent and role limit access
Author lets admins and owners build and test without publishing. Members can change accessible tasks and read published workflows and runs, never drafts; viewers can only read accessible tasks and projects.
A selection of tools. Each level includes the ones before it. The MCP docs list every tool. These workflow tools require admin/owner for OAuth. Publish consent can delegate approvals and coding-agent replies wherever you could answer in the console, even when API-key decisions or replies are disabled. Kovca does not ask for a new confirmation before each authorized MCP call.
Prefer a workspace API key?
For clients that send a bearer header, create an Author key in Settings → API keys and replace the placeholder below. Keep it secret. API keys still need each approval or coding-agent step's API opt-in.
Browser connections can be disconnected across workspaces in Account settings. Ordinary sign-out keeps them; Sign out everywhere revokes them. The kovca CLI keeps its existing device login.
Powerful for the agent, safe for you
The MCP server calls the same functions as the Studio and the API. It's never a second, looser way to do things.
Every edit is visible
Agents edit the same drafts as the Studio, through the same validated operations, so you can watch, undo, or take over at any point.
Credentials stay sealed
Agents never see app tokens or secrets. Connecting an app returns a link for a person to approve on the Connections page.
Access you choose
Choose Run, Author, or Publish. Browser sign-in also limits the client to your role and project permissions in one workspace; Author cannot publish.
Workflows become tools
Expose a published workflow as its own wf_ tool, with its input schema and a description, so any agent can run it like a function.
Kovca's agent, in your terminal
Run kovca agent and say what should happen. It picks the nodes, wires them up, validates, and tests on your examples, showing every step as it goes, and asks before it publishes.
- Tests use real Decision AI and Generative AI, while HTTP and app actions are simulated.
- In print mode (-p) it stops with exit code 7 before publishing, until you confirm or pass --yes.
- Your conversations are shared with the agent in the console.
See what the agent built, without a browser
kovca studio shows a workflow as a live graph with each node's settings beside it. Test it and watch the steps light up as they run, then publish.
- arrows
- walk the graph
- enter
- edit a setting
- a
- add a node
- T
- test the draft
- p
- publish
- /
- ask the agent
- shift+tab
- Agent mode
Every action is a command
Set KOVCA_API_KEY and the same commands run in CI: data on stdout, progress on stderr, and exit codes you can branch on.
Same rules as the web
The CLI edits the same drafts, runs the same validation, and publishes the same immutable versions as the Studio and the API.
Test before anything ships
kovca test runs the draft like the Studio's Test: AI steps are real, and actions are simulated unless you pass --live-actions.
Workflows as code
kovca workflows export and apply keep a workflow in your repository, so changes go through pull requests and CI.
An agent on every page of the console
Press Ctrl+J anywhere. It knows the page you're on, sets up projects, tasks, and workflows from one description, and wires them together: a task moving to Review starts a workflow, and the workflow writes back to the task.
For a bigger request it proposes a plan first and builds nothing until you approve it. It can only do what you could do yourself, and it never touches app connections, members, API keys, or billing. How Kovca's agent works
Workflows can hand work to a coding agent
A coding agent step starts GitHub Copilot, Cursor, or Claude on your repository with your own account, and the run waits while it works. If the agent asks a question, a person answers it from Approvals, the run page, or the CLI; then the run continues with its pull request.
Prefer to see it on a canvas?
Everything an agent builds opens in the Studio, where you can inspect, test, and change it by hand.