Connect your agent through MCP

Add Kavela to Claude Code, Codex, Cursor, VS Code, Claude or any MCP client, so your agent can save a session as a workflow that runs in the cloud, check how it runs, and change it.

Kavela's MCP server lets an agent you already use work with your Kavela workspace. From the agent's chat you can save a session as a workflow, run it, read how its runs went, change it, hand Kav new work, and answer decisions waiting for you. The workflows run in Kavela's cloud, so your computer can be off.

Setting	Value
Server address	`https://mcp.kavela.ai/mcp`
Sign-in	Your Kavela account, in the browser. No key to paste
Transport	Streamable HTTP
Without a browser sign-in	A Kavela API key, sent as `Authorization: Bearer <key>`

Problems this solves

Problem	What Kavela does
A session got the job right, and it's gone when the chat closes	Your agent saves it as a workflow, with what to deliver and what it must never do, on a schedule you set
An automation stops working and nobody notices	A reviewer checks every run. Ask your agent how a workflow is doing and it reads you the passes, the misses and the check missed most
Changing an automation means rebuilding it	Say what should change. It is saved as a new version, and the next run uses it
The job only runs while your laptop is open	Workflows run on Kavela's agents in the cloud

Claude Code

Add the server:

claude mcp add --transport http kavela https://mcp.kavela.ai/mcp

Then sign in. In Claude Code, run /mcp, select kavela and choose Authenticate. A browser window opens for your Kavela sign-in.

Codex

Add the server:

codex mcp add kavela --url https://mcp.kavela.ai/mcp

Then sign in:

codex mcp login kavela

Or add it to ~/.codex/config.toml yourself:

[mcp_servers.kavela]
url = "https://mcp.kavela.ai/mcp"

Cursor

Use the Add to Cursor button on kavela.ai/mcp, or add this to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "kavela": {
      "type": "http",
      "url": "https://mcp.kavela.ai/mcp"
    }
  }
}

Then open Cursor's MCP settings, click Needs login next to kavela, and sign in.

VS Code

Use the Add to VS Code button on kavela.ai/mcp, or add this to .vscode/mcp.json:

{
  "servers": {
    "kavela": {
      "type": "http",
      "url": "https://mcp.kavela.ai/mcp"
    }
  }
}

Then allow the Kavela sign-in when VS Code asks, and sign in.

Claude

In Claude on the desktop or the web:

Open Settings, then Connectors, then Add custom connector.
Name it Kavela and paste https://mcp.kavela.ai/mcp.
Click Add, then Connect, and sign in.

Other clients

Client	How it connects
OpenCode	Add a remote server at the address in `opencode.json`, then run `opencode mcp auth kavela`
Windsurf, Cline, Zed	These speak local stdio. Use the `npx -y mcp-remote https://mcp.kavela.ai/mcp` bridge, which carries the browser sign-in on first run
Any other MCP client	Add an HTTP MCP server at the address. If it can't sign in through a browser, send a Kavela API key as a Bearer header

To create an API key, open Settings in Kavela, then API keys. A read-only key can list and read, and changes nothing.

Check it works

In Claude Code, run claude mcp list. kavela shows as connected.
Ask your agent: "Call start_here." It replies with your active workspace, its agents, its workflows and how many decisions are waiting for you.

What your agent can do

Group	Tools	What they do
Start	`start_here`	Shows the workspace and which tool to use for what
Kav	`chat_with_kav`	Hands Kav, the workspace manager, any request or follow-up, the same as typing it in Kavela
Workflows	`create_workflow`, `list_workflows`, `read_workflow`, `update_workflow`, `delete_workflow`	Save a workflow, see them all, read one in full, change or pause one, delete one
Versions	`list_workflow_versions`, `restore_workflow_version`	See every version and how its runs went, bring an earlier one back
Runs	`run_workflow`, `workflow_status`	Start one run now, read the latest runs and what they missed
Files	`list_files`, `read_file`, `write_file`, `edit_file`, `delete_file`, `search_workspace`, `upload_to_workspace`, `download_from_workspace`	Work with the workspace's files
Decisions	`list_approvals`, `respond_approval`	See the decisions waiting for you and answer one
Workspaces	`list_workspaces`, `switch_workspace`	Pick the workspace the tools act on

Everything your agent starts through MCP also shows in Kavela: the same chat, workflow and files.

Save a workflow

A workflow is an instruction Kav follows on every run, picking the agents for each one. Your agent writes the instruction from your session. A good one says what to deliver and anything the run must never do.

How it starts	What to say
On a schedule	Times of day and days, such as weekdays at 08:00, or a gap such as every 30 minutes. Times are in your own time zone
When something happens	It gets its own web address that another service sends data to. The signing secret is shown once, to you, in Kavela, never through MCP
When you run it	It runs only when someone runs it

Change a workflow

You want	Say
A different instruction	"Change the morning digest to cover 5 posts."
A checklist	"Add a check: every post has its link." A checklist holds up to 12 lines. Mark a line as must pass when a run should fail without it
A pause	"Pause the morning digest." It stops starting on its own, and `run_workflow` still runs it
An earlier version	"Put it back to version 1." The earlier version is copied as the new live one, so nothing in the history is lost

Each change to the instruction or the checklist is a new version, and the next run uses it. If someone changed the workflow in Kavela since your agent read it, nothing is saved and your agent is told the new version number, so a change is never overwritten. Kavela doesn't call one version better than another until each has 10 checked runs.

First prompts

Ask	What happens
"Save this session as a workflow that runs every weekday at 8."	`create_workflow` saves it on a schedule
"Run it once now."	`run_workflow` starts one real run in the cloud
"How did the morning digest do this week?"	`workflow_status` reads the latest runs, the pass rate per version and the check missed most
"Is anything waiting on me?"	`list_approvals` lists the decisions waiting for you

What stays in Kavela

A final sign-off, a password or key, and taking over a live browser can only be answered by you in Kavela. Your agent can see that they are waiting, and Kavela shows them in your Approvals.

If something goes wrong

What you see	What to do
The agent can't reach Kavela, or asks you to sign in again	Sign in again: in Claude Code, run `/mcp` and authenticate `kavela`
The tools act on the wrong workspace	Ask your agent to call `list_workspaces`, then `switch_workspace`
"This key is read-only"	The key can list and read only. Create a key that can change things in Settings, then API keys
"This workspace has no Kav yet"	Open the workspace in Kavela once, then try again

Connect your agent through MCP

On this page