Short answer: LumisReach is a Rush partner and is available as an MCP server on Rush, the platform that runs Claude Code, Codex or OpenCode in their own cloud computers. Add the LumisReach MCP server at https://mcp.lumisreach.com/mcp as a Rush connection with your LumisReach API key as a bearer token, name it on a run or a saved agent, and that agent can place AI phone calls, run AI text conversations, read transcripts and buy numbers while it works. Setup takes two requests and about five minutes.
Rush details below come from Rush's public docs and pricing page, checked November 19, 2026. LumisReach prices are our published rates on the same date.
What is Rush, and why run a phone-calling agent there?
Rush is a platform from Phoenix that runs coding and work agents in the cloud. Each run gets its own cloud computer, shared with no other run, and comes back with the result plus a record of every step it took and what it cost. A run is started from the console, from one API request, from an MCP client, or by an automation on a schedule or an event.
That shape fits phone work better than a chat window does. An AI phone call takes minutes: the agent dials, maybe retries, maybe leaves a voicemail, and the outcome arrives later. Many MCP chat clients give a tool call about a minute before they give up. A Rush run is meant to take minutes; Rush's own MCP docs say so plainly in a section titled "A run takes minutes". An agent in a Rush run can start a call, keep working, and poll the LumisReach job until it settles, and when a turn ends the run parks with its memory and files intact at no cost until you write back.
Why an MCP tool instead of letting the agent drive a dialer?
A Rush run comes with a browser and a desktop, so in principle an agent could log in to a calling app and click through it. Rush's own engineering posts explain why that is the expensive path. In Your Agent Can Use Any App. For a Quarter Fewer Tokens. the team starts from the premise that "Most apps have no API. Most have no MCP server either," and measures how much every screenshot costs an agent that has to work through a screen. In The Other Way to Speed Up a Browser Agent they describe actions dropping from 30 to 60 seconds in a live browser to roughly 200 milliseconds once a typed client existed for the same work.
Phone calls are a case where the typed client already exists. One start_ai_phone_call tool call
replaces the whole loop of finding the dialer, typing the number and watching the screen, and the
voice conversation itself runs on LumisReach's telephony stack, not inside the agent's context
window. The agent spends its tokens on deciding who to call and what to ask, and reads back a short
outcome and summary.
How do you connect LumisReach to Rush?
Rush calls a tool an agent can use a connection: an MCP server, or a built-in connector such as Gmail or Slack. You sign in once and every run names the connections it wants. LumisReach is a Rush partner and is listed on Rush as an MCP server, next to the apps in Rush's integrations directory, so phone calls and texts sit alongside Gmail, Slack and the other tools a run can use.
- Create a LumisReach key. In the LumisReach app, open Settings, then Developer API. The key starts with
lumisreach_, is shown once, and works on the whole v1 API and every MCP tool. Add a card under Billing for anything that costs money; reads are free. - Add LumisReach as a Rush connection. In the Rush console, open Settings, then Connections, and add an MCP server with the URL
https://mcp.lumisreach.com/mcpand bearer auth. Or do it from the API withPOST /me/connections(below). - Name it on a run. Put
"lumisreach"in the run'sconnectionslist, or save it on an agent once. - Give the agent a task. Describe the outcome in plain words; the agent picks the LumisReach tools itself.
- Read the result. The Rush session holds the agent's steps and cost; the LumisReach workspace holds the call recording, transcript, summary and contact.
Set the auth mode explicitly. The LumisReach MCP server takes the same bearer API key as the REST
API and has no OAuth sign-in, and Rush's docs say that when auth is left out Rush probes the server
and falls back to a sign-in if it does not answer, which is the wrong path here.
curl -X POST "https://api.phoenix.so/me/connections" \
-H "Authorization: Bearer $RUSH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"type": "mcp",
"name": "lumisreach",
"url": "https://mcp.lumisreach.com/mcp",
"auth": {"mode": "bearer", "token": "'"$LUMISREACH_API_KEY"'"}
}'
Rush stores the token and never returns it; the connection reads back with its status and the list
of tools it found. From then on, any run of yours can name lumisreach.
What does a full request look like?
A run that names the connection and hands the agent a phone task:
curl -X POST "https://api.phoenix.so/o/$RUSH_ORG/sessions" \
-H "Authorization: Bearer $RUSH_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"agent": "claude",
"input": "Call Dana at +1 415 555 0123 and ask whether she still wants the roof quote and when she is free for a site visit. Try twice if she does not pick up. Report the outcome and a one-line summary of the call.",
"connections": ["lumisreach"]
}'
You can also skip the saved connection and pass the server inline for one run and its follow-ups, as Rush's files, skills and connections lesson shows. The token is stored encrypted on that run and never saved to your connections:
"connections": [
{"url": "https://mcp.lumisreach.com/mcp", "name": "lumisreach",
"auth": {"mode": "bearer", "token": "lumisreach_…"}}
]
Inside the run, the agent does what it would do in any MCP client:
start_ai_phone_call → job queued, max_attempts 2
get_quick_action_status → in_progress
get_quick_action_status → completed, outcome goal_achieved
Agent: Dana picked up on the second attempt. She still wants the quote and can do Thursday after 3pm for the site visit.
The call, its recording and transcript, and Dana's contact record land in the LumisReach workspace like any other conversation, so the people on your team see exactly what the agent did.
Should you save it as an agent?
If the same kind of phone work comes up again, yes. A Rush saved agent keeps the harness, model, standing instructions, skills and connections, and every run names only the agent and the task. Save one with LumisReach on it and with instructions that encode how your business talks to people:
{"name": "Callback agent", "harness": "claude",
"instructions": "You follow up with leads for a roofing company by phone and text using the LumisReach tools. Prefer one AI call with up to 2 attempts; if it ends in voicemail, send one short text. Never call before 9am or after 7pm in the lead's time zone. Report each lead's outcome in a table.",
"connections": ["lumisreach"]}
That saved agent is also what a Rush automation can run, which is how a daily callback list or a "call every new form lead" webhook works. We cover that pattern in scheduled AI calls and texts with Rush automations.
Which LumisReach tools will the agent see?
The connection exposes the same MCP tools any client gets. The ones an agent reaches for most in a Rush run:
| Tool | What it does |
|---|---|
start_ai_phone_call | An AI call toward a goal, with retries and voicemail handling. No number of your own needed |
start_ai_text_conversation | An AI text thread that keeps answering replies toward a goal |
send_quick_text / send_quick_email | One message sent exactly as written |
get_quick_action_status | The job's per-person status, outcome and summary |
list_calls, get_call, get_call_recording | Call history, speaker-labelled transcripts and audio |
list_contacts, get_contact, list_messages | Everyone you have reached and the whole thread with each |
setup_ai_phone_number, buy_phone_number | A number of your own with an AI that answers it |
ask_lincoln | Hand a whole job, such as a campaign, to LumisReach's own agent |
The full list and every parameter are in the MCP server reference.
What does it cost?
Two bills, each for its own half of the work.
| Item | Price |
|---|---|
| Rush Free | $1 one-time credit that covers the AI model and the cloud computer, 1 run at a time |
| Rush Pro | $40 per member a month: 40 cloud computer hours, 4 runs at once, bring your own model |
| Rush Max | $300 per member a month: 200 cloud computer hours, 12 runs at once, bring your own model |
| LumisReach quick AI call | $0.20 per call placed (including retries) plus $0.20 per minute, billed by the second |
| LumisReach quick AI text | $0.05 per outbound message; replies are free |
| LumisReach reads (history, transcripts, contacts) | Free |
| LumisReach API and MCP only plan | $20 per month, credited back as $20 of usage |
Rush counts computer hours only while a run's computer is awake, so a run parked waiting for your reply costs nothing there (Rush pricing). Runs started with a Rush API key or from an MCP client use Rush's models and are paid from prepaid credits; Rush's usage and billing page has the details. Our side is on the LumisReach pricing page.
A worked example: a run that calls 10 leads, 7 of whom talk for about 2 minutes, costs on the LumisReach side 10 × $0.20 plus 14 minutes × $0.20, about $4.80, plus the Rush computer time and model tokens the agent used to plan and report.
What are the limits?
- Bearer key, not OAuth. The LumisReach MCP server authenticates with an API key. Rush's per-user agents, where each of your app's users connects their own accounts, accept only MCP servers that sign in with OAuth and refuse token-based ones, so LumisReach works on your own runs and saved agents but not on agents made from a template for your end users.
- Connections are personal. A Rush connection belongs to the person who added it, and a run started with a Rush API key uses the connections of the member who created that key. The LumisReach key decides which workspace is billed and whose contacts the agent sees.
- Quick calls spend straight away.
start_ai_phone_calland the quick text tools charge when they run. Approval prompts exist only when the agent hands work to Lincoln throughask_lincoln, and only if a person, not the agent, answers them. Our post on trusting an agent with your phone line covers how to keep a person in that loop. - Shared senders by default. Quick calls and texts go out from LumisReach's registered shared senders, so callbacks do not reach a line of yours; buy a number for $5 a month when people need to call you back.
Frequently asked questions
Is LumisReach available as an MCP server on Rush?
Yes. LumisReach is a Rush partner and its MCP server at https://mcp.lumisreach.com/mcp is listed on
Rush as a connection any run can use. Add it once in Rush under Settings, Connections (or with
POST /me/connections) using bearer auth and your LumisReach API key, then name it in a run's
connections or on a saved agent.
Can a Claude Code agent on Rush make real phone calls?
Yes. With the LumisReach connection named on the run, Claude Code, Codex or OpenCode can call the
start_ai_phone_call tool with a phone number and a goal. LumisReach places the AI call, retries if
nobody answers, handles voicemail, and returns the outcome, summary and transcript, which the agent
reads with get_quick_action_status and get_call.
Do I need my own phone number for a Rush agent to call or text?
No. Quick AI calls and texts go out from LumisReach's registered shared senders, so a Rush agent can reach someone minutes after you create the key. Buy your own number for $5 per month when people need to call you back or volume outgrows the shared pool.
Why does Rush need the auth mode set to bearer?
Because the LumisReach MCP server uses API-key bearer auth, not OAuth. When a Rush connection is
added without an auth field, Rush probes the server and starts a browser sign-in if it does not
answer, so pass {"mode": "bearer", "token": "lumisreach_…"} explicitly.
What does a phone call from a Rush agent cost?
On LumisReach, $0.20 per call placed plus $0.20 per minute; a text is $0.05. On Rush, the run uses cloud computer time (free while parked) and model tokens, covered by a $1 free credit, prepaid credits, or a Pro or Max plan depending on how the run was started.
