AI assistant (MCP)
Connect Parallel to Claude — or any MCP client — and ask it to do things in your profiles. “Open my Reddit profile and check for new messages.” It opens the profile, navigates, reads the page, clicks, types, and can take a screenshot to see what it is doing.
The important part: it never handles your logins. The steps run inside Parallel, through the same engine the RPA editor uses, so every action arrives with the profile’s own fingerprint, proxy and signed-in session already in place. The assistant does not know those exist and cannot get them wrong.
What it can do
Anything the RPA editor can do, because it is the same engine underneath. In practice that means: open and close profiles, go to a page, read its text, click things, fill fields, take screenshots, and run one of the automations you have already built and tested.
It is a local connection. The assistant talks to Parallel on your own machine, not to a server, and Parallel has to be running for any of it to work.
Setting it up
Open API & MCP in Parallel’s sidebar. From v0.0.55 there is a Set up for Claude Code… button: pick the folder you open with Claude Code and Parallel writes the config file for you, with the path, port and token already filled in, keeping any other MCP servers you have configured there.
On earlier versions, click Copy config and save it yourself as a file called
.mcp.json in that folder. It is a leading-dot filename, which Windows Explorer resists
creating — easiest is to save it from a text editor with the quotes around the filename, or to
update Parallel and use the button.
Then open that folder in Claude Code — a new session, since servers are found when one starts — and ask it to list your profiles. If it names them, you are connected.
There is a Copy config button too, if you would rather place the file yourself.
Claude Desktop
Depends on the build. If your Claude Desktop has Settings → Developer → Edit Config, paste the entry into the file it opens and restart the app completely, tray icon included.
If it does not have that option, that build cannot host a local MCP server — its Connectors screen
takes remote URLs only, and editing claude_desktop_config.json by hand does not work,
because the app rewrites that file from its own settings and an entry added by hand disappears
within minutes. Use Claude Code instead; it is the same server and the same tools.
The tools
Twelve, in three groups.
Profiles — list_profiles, start_profile,
stop_profile, profile_status.
Driving the page — navigate, read_page, click,
type_text, screenshot. These start the profile if it is not already running,
so there is no need to open it first.
Your automations — list_processes, run_process,
run_steps. Prefer running a process you have already built and tested over asking the
assistant to reconstruct the same steps each time: it is faster and far less likely to go sideways.
A step that fails comes back as a readable result rather than an error, so the assistant can look at the page and try something else instead of repeating itself.
Seeing what it did
Everything an assistant does is recorded in RPA → Task Log, tagged assistant and named by its steps. A Started by filter separates those runs from ones you started yourself, and Log detail opens the step-by-step trace.
This is worth knowing about before you need it. A run started by something other than you is the one you are most likely to go looking for afterwards.
Before you start
An assistant driving a profile is acting as you. It is signed into whatever that profile is signed into, and there is no line between reading a page and clicking Send. Start it on a profile you do not mind it fumbling, and watch what it does the first few times.
The token is the whole key. The config file holds a token that can open any profile in your install, which means acting as every account they are signed into. Treat it like a password, keep the file out of anything you publish, and use Regenerate token in API & MCP if it ever leaks — the old one stops working immediately.