Agents and MCP

Local wb mcp

wb mcp serves Workbench tools to a coding agent over stdio, from the CLI on your machine. It runs, supervises, and lists local work and local projects only. For cloud work, connect the agent to mcp.workbenches.dev; see Connect an agent.

Connect#

Claude Code:

claude mcp add wb -- wb mcp

Any other MCP client, with the stdio configuration:

{
    "mcpServers": {
        "wb": {
            "command": "wb",
            "args": ["mcp"]
        }
    }
}

Tools#

wb mcp provides 13 tools: search, view, run, wait, send, answer, cancel, runs, sessions, push, projects, create_project, and tasks. They use the same catalog, inspection, supervision, control, history, and registry operations as their CLI counterparts.

Tool	Does
`search`	Always includes saved Workbenches, and adds registry results when the CLI already has a registry login.
`view`	Inspect a Workbench, as `wb view`.
`run`	Start a detached local run from a saved alias. Accepts `project`, `dir`, and `runtime` to select a declared `local`, `docker`, `e2b`, or `daytona` runtime; without `runtime` it uses the Workbench's default.
`wait`	Return a cursor at a turn boundary, pending input, timeout, or terminal state. Each call is capped at 60 seconds.
`send`, `answer`, `cancel`	Control a local session, as `wb send`, `wb answer`, and `wb kill`.
`runs`, `sessions`	Local run and session history.
`projects`, `create_project`, `tasks`	Local projects and their tracked tasks.
`push`	Push to the registry. Never starts a login flow.

Local runs refuse to start while a worker from an earlier version of wb is still active, as wb run does.

The loop#

Find a saved or published Workbench with search, inspect it with view, then start it with run. Then run, wait, send or answer, wait again. wait returns early with a cursor when a turn ends, something needs input, or 60 seconds pass. A run that is still going is normal: call wait again with the returned next until the state is completed, failed, or cancelled.

How it differs from mcp.workbenches.dev#

	`wb mcp`	`mcp.workbenches.dev`
Where it runs	On your machine, as a stdio subprocess	Hosted, Streamable HTTP
Where runs execute	Your machine: `local`, `docker`, `e2b`, or `daytona`	Cloud sandboxes, or your organization's workers
Credentials	Your local `wb` setup: `wb connect`, environment, runtime keys	An organization API key or OAuth
What `run` takes	A saved alias, plus `project`, `dir`, `runtime`	`org/name` or a `version_id`, plus `repository`, `ref`, `branch`, `worker`, `max_minutes`
`wait`	At most 60 seconds per call; returns at turn boundaries	Up to 3600 seconds per call; returns only when a run needs you or ends
Projects	Local projects in `wb.db`	Organization projects on workbenches.dev
Extra tools	`tasks`	`workers`, `changes`
Login	Not required. Registry search and push use an existing `wb login`	Required

Local sessions also show in wb status and, when tracked in a project, in wb studio. See Projects, status, studio, and MCP.