One context packet per turn
Who the visitor is, what they asked before, the page they are on, open tasks and your instructions, in one typed JSON object. No scraping a UI.
Plug in Claude, GPT, Grok or your own agent. Relaydesk gives it the inbox, the widget, the tools and the guardrails to run your customer support, and pulls in a person when it should.
14-day free trial · No card · Bring your own model key
New visitor
via chat widget
Bring the agent you already trust
Built for agents, not people clicking buttons
Every support tool you know was designed for humans in an inbox, with an AI bolted on and billed per answer. Relaydesk is the other way around. Your agent gets typed context, real tools, a knowledge base and an audit trail. Your team gets the hard conversations, and a bill that does not grow every time a customer says hi.
Connect in minutes
Run a hosted agent on your own model key, plug in the agent you built, or let an MCP client work the inbox. Mix them, route between them, swap them any time.
1POST /v1/agents2{3 "name": "Frontline",4 "provider": "anthropic",5 "model": "claude-sonnet-4-5",6 "instructions": "Cite the docs. Hand off when unsure.",7 "guardrails": { "monthlyCostCapUsd": 50, "redactPii": true }8}Every core feature, agent-shaped
Who the visitor is, what they asked before, the page they are on, open tasks and your instructions, in one typed JSON object. No scraping a UI.
Point the agent at your own HTTP endpoints. Mark the risky ones as needs-approval and a teammate taps approve.
your_risky_tool
waiting for approval
Docs, sitemaps, PDFs and Q&A. Answers cite their sources, so you can see why it said what it said.
Keywords, routing rules or the agent's own judgment send a chat to a person, with the full trace attached.
Every model call, tool call and guardrail, step by step. Replay conversations against a new prompt before you ship it.
Monthly cost caps on your own key, turn limits and PII redaction. Your agent cannot run up a surprise bill.
About 21 KB, shadow DOM, light and dark, rich blocks, CSAT. Visitors can attach files and voice notes.
Images, PDFs, Office files, spreadsheets, audio and video become model-ready input automatically.
Signed webhooks, server-sent events and an MCP server. Email alerts for your team when a person is needed.
How it works
From the widget on your site or app. Identity is verified with a signed hash, so the agent knows who it is talking to.
Customer, history, page, attachments and your instructions become one context packet, ready for any model.
Hosted on your key, on your server via webhook, or over MCP. It holds a lock, so two agents never answer at once.
Lookups run straight away. Anything you marked as risky waits for a teammate to approve, in the dashboard or by email.
Most chats close on their own with a CSAT check. The rest land with a person, with the whole trace attached.
$0
per resolution, ever
0%
markup on model tokens
~21 KB
widget, gzipped
14 days
free trial, no card
Pricing
Try any plan free for 14 days, no card. Then one flat price: no seats, no per-resolution fees, no token markup, no overages. Your model provider bills you for tokens at cost, and we never touch that line.
One site, one brand, an agent that never sleeps.
billed yearly, save 17%
Start free trialMore volume, more brands. Still one flat price.
billed yearly, save 16%
Start free trialFor companies running support on agents across brands.
billed yearly, save 17%
Start free trialYour cloud or ours, with the paperwork done.
volume and contract terms
Talk to usEvery plan starts with a 14-day free trial, no card needed. A conversation is one visitor thread that gets a reply in a calendar month, however many messages, agents or tool calls it takes. Going over your band never costs extra; if you are over two months in a row, we will suggest the next plan. Prices in USD, before tax.
Do the math
Compared with Intercom's cheapest plan plus its AI agent. Both on yearly billing.
$24 plan + about $160 paid to your model provider
5 × $29 seats + 2,000 × $0.99 resolutions
You keep
$1,941 / month
Estimates. Model cost assumes about $0.08 per conversation on a Sonnet-class model and varies with your model and chat length. Intercom figures use public list prices checked November 2026 and assume every AI chat counts as a resolution.
FAQ
You add your own Anthropic, OpenAI or xAI key (or run your own agent). Model usage is billed by your provider at their price. Relaydesk charges one flat plan price and never marks up tokens or charges per answer.
Every plan starts with 14 days free, no card, up to 1,000 conversations. When it ends, your widget switches to a leave-your-email form instead of going blank, and your data is kept for 60 days so you can pick up where you left off.
One visitor thread that gets at least one reply in a calendar month, however many messages, agents or tool calls it takes. Plans come in bands, not meters: going over never costs extra, and if you are over two months in a row we suggest the next plan.
Because your agent does the work. Charging per seat punishes you for adding a teammate to review a few chats, and charging per resolution punishes you for the agent doing its job. Flat plans keep the bill boring.
It depends on the model and how long your chats are. A typical support conversation on a mid-tier model costs a few cents. You can set a monthly spend cap per agent so it never surprises you.
No. Nothing about your business is hardcoded. You decide which tools exist, which need a person's approval, what the knowledge base contains and when to hand off.
Yes. Invite your team with owner, admin or agent roles. They get email alerts when a chat needs them and can take over any conversation from the inbox.
Relaydesk ships as Docker images for the API, the agent runner, the MCP server and the dashboard. Talk to us about Enterprise if you want to run it in your own cloud.
Set up the widget, connect a model and answer your first chat in about ten minutes. Free for 14 days, one flat price after that.