What’s the best infrastructure for AI agents in 2026?
By NoClick, checked November 8, 2026. We make NoClick, listed first; every other company’s facts and prices are from its own pages, linked beside them.
The short answer
NoClick covers every layer an agent product needs, under one key and one bill. To assemble specialists instead, these are the picks per layer.
- Every layer under one key and one billNoClick (we make it): agents, sandboxes, browsers, inboxes, numbers, connections and MCP servers, each at cost
- Run an always-on agentCloudflare Agents, or Amazon Bedrock AgentCore if you’re on AWS
- Run the code it writesE2B or Daytona, or boat for a whole persistent VM
- Give it a browserBrowserbase or Steel
- Give it an inboxAgentMail
- Give it a phone numberTwilio or Telnyx, with Vapi or Retell AI for the voice
- Act in your users’ accountsComposio or Nango
- Host its MCP serverCloudflare Workers or Vercel
NoClick.Every layer, one key.
NoClick
At what each service costsOne SDK, API and MCP server for every layer on this page: hosted agents, apps, sandboxes, browsers, inboxes, phone numbers, your users’ connected accounts and spending caps, each at what it costs NoClick.
- Best for
- A product that needs several of these layers for each of its users, under one key, one bill and per-user spending caps.
- Falls short
- Each layer covers what most agent products need; for specialist needs such as GPU sandboxes, a dedicated provider goes further.
How it works and full pricing
An agent made with the SDK keeps its conversations and a memory of named notes, and NoClick runs its loop (or Claude Code, Codex, OpenCode, Hermes Agent or OpenClaw in a sandbox), waking it at a time or when work finishes. Sandbox.create runs a Python or Node image; Browser.create gives a browser that keeps a person’s sign-ins; Inbox.create makes an address whose mail can wake an agent; a bought number sends every call to your agent through a live voice; Connection.from_name gives each of your users a connect page; and McpServer.from_name puts up to 64 tools at one URL. Each of your users can have a monthly cap, and an agent’s purchases wait for the account owner’s yes.
Every service at what it costs NoClick, nothing added: an inbox is free and a reply $0.00035; a US number $1 a month and $0.0607 a minute answered, voice included; a browser $0.48 an hour; a sandbox $0.1419 per core-hour; hosted MCP servers free. Free includes $1 a day of usage, up to $5 a month, with a 15% fee on funds added; Member is $25 a month with $10 included and a 5% fee.
At a glance.
We started from the products five AI chatbots named when we asked them these questions on November 8, 2026, added leaders they missed, and compared what each does for an agent, how it bills, its free tier and its limits, from each company’s own pages read the same day.
| Layer | NoClick | Top picks | From |
|---|---|---|---|
| Running the agent | Yes: Hosted agents with memory | Cloudflare Agents, Bedrock AgentCore | $2.19 a month (Fly.io) |
| Code sandboxes | Yes: Python and Node sandboxes | boat, E2B, Daytona | $0.036 an hour for 4 vCPU (boat) |
| Cloud browsers | Yes: Keeps a person’s sign-ins | Browserbase, Steel | $0.02 an hour (Browser Use) |
| Email inboxes | Yes: Free inboxes, $0.00035 a reply | AgentMail, Nylas | 3 inboxes free |
| Phone numbers, SMS and voice | Yes: $1 a month, voice included | Twilio, Telnyx; Vapi, Retell AI | $0.50 a month a number (Plivo) |
| Connected accounts | Yes: Connect links, tokens kept fresh | Composio, Nango | 100,000 calls free (Composio) |
| MCP server hosting | Yes: Free hosted MCP servers | Cloudflare Workers, Vercel | Free tiers |
| Spending and payments | Yes: Caps and owner approvals | Stripe Issuing | $0.10 a virtual card |
Running the agent
Keeps the agent’s process up, routes messages to it, and keeps its state and memory between conversations.
NoClick runs hosted agents that keep their memory between conversations and wake at a time or when work finishes. Agents in the docs
Cloudflare Agents (Durable Objects)by Cloudflare
From $5 a monthAn SDK where each agent is a Durable Object with its own SQL storage, WebSockets and scheduled work.
- Best for
- Many small stateful agents (one per user, room or document) that sit idle most of the time.
- Falls short
- You build for Cloudflare’s Workers runtime and its SDK; heavy or long native workloads need Containers or another host.
How it works and full pricing
Every agent gets a durable identity and local SQL storage, holds real-time connections and schedules its own work. Objects that are idle and eligible for hibernation are not billed for duration, so an agent waiting on its next message costs little.
Workers Paid: $5 a month minimum. Durable Objects include 1 million requests and 400,000 GB-s a month, then $0.15 per million requests and $12.50 per million GB-s; SQLite storage $0.20 per GB-month past 5 GB-month. Workers Free allows SQLite Durable Objects at 100,000 requests a day.
Sources:developers.cloudflare.comdevelopers.cloudflare.comdevelopers.cloudflare.com
Amazon Bedrock AgentCoreby Amazon Web Services
From $0.0895 per vCPU-hourAWS’s managed runtime for agents and tools, with separate Memory, Gateway, Browser and Code Interpreter services.
- Best for
- Teams already on AWS who want per-session isolation and IAM-based identity.
- Falls short
- Runtime, memory, gateway and tools are metered separately, so a month’s bill takes work to predict, and setup assumes AWS accounts and IAM.
How it works and full pricing
Each user session runs in its own microVM, for up to 8 hours (up to 14 days on Instances), with any framework or model, over HTTP, WebSockets, MCP or A2A. CPU isn’t billed while the agent waits on I/O such as a model’s reply; AgentCore Memory keeps short- and long-term memory as a separate service.
Runtime v1: $0.0895 per vCPU-hour and $0.00945 per GB-hour; v2: $0.1276 per vCPU-hour and $0.0169 per GB-hour. Memory: long-term $0.75 per 1,000 records a month; short-term $1.00 per GB ingested from October 6, 2026. New AWS customers get up to $200 in Free Tier credits.
LangSmith Deploymentby LangChain
$39 per seat a monthLangChain’s managed hosting for LangGraph agents, with a database that persists their state.
- Best for
- Teams that already build with LangGraph and use LangSmith.
- Falls short
- It’s built around LangGraph, seats are paid per developer, and hybrid or self-hosted deployment is Enterprise only.
How it works and full pricing
You deploy a LangGraph agent and LangChain runs it with a database that keeps its state live; uptime is the time that database is live and persisting state. It sits beside LangSmith’s tracing and evals.
Plus: $39 per seat a month, with 1 free small serverless deployment. More is metered in LSUs ($1 each): runtime 0.0675 LSU per vCPU-hour and 0.0090 per GiB-hour, database 0.177 LSU per vCPU-hour and 0.025 per GiB-hour. The free Developer plan has no deployments.
Sources:langchain.com
Small VMs from your Docker image, billed by the second, in regions you pick.
- Best for
- A cheap always-on process for an agent you’ve written yourself.
- Falls short
- Plain VMs: restarts, state, memory and scaling are your code’s job.
How it works and full pricing
You run the agent as a long-lived process on a Machine and keep its state on a volume or in a database you add. Machines can stop or suspend when idle and wake on the next request.
No plan fee. An always-on shared-cpu-1x Machine with 256 MB is $2.19 a month (cheapest regions); volumes $0.15 per GB-month, billed on provisioned size. New users get a $1 trial.
Sources:fly.io
A platform that runs services from a repo or image, with databases and volumes beside them.
- Best for
- Getting an agent service and its database running together with little setup.
- Falls short
- Nothing agent-specific: memory, scheduling and isolation between users are yours to build, and always-on compute is billed for every second.
How it works and full pricing
You deploy the agent as an always-on service and add Postgres or a volume for what it remembers. Usage is billed per second on top of the plan fee.
Hobby: $5 a month including $5 of usage; Pro: $20 a month per workspace including $20. Usage about $20 per vCPU a month, $10 per GB of RAM a month, volumes about $0.15 per GB-month. Free plan: $1 a month of usage.
Sources:railway.com
Code sandboxes
Runs code the model wrote, or a user’s upload, in an isolated machine away from your servers.
NoClick’s sandboxes run Python or Node with volumes, secrets and limited networking. Sandboxes in the docs
Persistent Linux VMs for agents, with SSH, Docker and a desktop, that fork from a snapshot.
- Best for
- Agents that need a whole persistent Linux machine (SSH, Docker, a desktop) at the lowest hourly price here.
- Falls short
- EU regions only for now (Germany, Finland and France), and a $20 monthly minimum.
How it works and full pricing
Each sandbox is an Ubuntu VM with its own IP address, SSH and SCP, Docker inside and a 60fps desktop, with Chrome, VS Code, Node.js, Bun and Rust installed, so Claude Code, Codex or OpenCode can run in it. boat stop snapshots it and stops the bill, and boat fork copies one at the disk level. It has an HTTP API, Python and TypeScript SDKs and a CLI.
Billed per second while running, with a $20 monthly minimum that includes 2,000,000 VM-seconds (about 555 hours of the default size). Per hour: small (2 vCPU, 4 GB) $0.018, default (4 vCPU, 8 GB) $0.036, large (8 vCPU, 16 GB) $0.072, xlarge (16 vCPU, 32 GB) $0.200. Saved snapshots are $0.056 a day. A 7-day free trial includes 25 compute hours.
Sources:boat.devdocs.boat.dev
Linux VM sandboxes created on demand for agents, with templates for what each starts with.
- Best for
- Agents that run many short pieces of code and need a sandbox per task.
- Falls short
- Sessions over an hour and more than 20 at once need Pro at $150 a month.
How it works and full pricing
Your code creates a sandbox, runs commands and moves files; a sandbox can be paused, keeping its filesystem and memory, and resumed later.
Usage on every plan: $0.000014 per vCPU-second and $0.0000045 per GiB-second; disk free. Hobby: $0 with $100 one-time credit, 1-hour sessions, 20 at once. Pro: $150 a month plus usage, 24-hour sessions, 100 at once.
Sources:e2b.devdocs.e2b.dev
Sandboxes on Modal’s serverless platform, with custom images, volumes and GPUs.
- Best for
- Sandboxes that need GPUs, large custom images or many hours.
- Falls short
- Per core-hour it’s the priciest here, and Team costs $250 a month.
How it works and full pricing
Sandboxes run on gVisor or a VM runtime, take custom images, volumes and GPUs, and live up to 24 hours; filesystem snapshots keep state longer.
Sandbox CPU $0.00003942 per core per second and memory $0.00000667 per GiB per second (3x Modal’s function rates). Starter: $0 plus compute, $30 a month free credits. Team: $250 a month plus compute, $100 free credits.
Sandboxes that are full computers with their own kernel, filesystem and network stack.
- Best for
- Long-lived agent workspaces at E2B-level prices with no plan fee.
- Falls short
- Capacity comes in tiers unlocked by top-ups (Tier 2 needs a $25 top-up for 100 vCPUs across the org), and per-sandbox limits aren’t on the public page.
How it works and full pricing
Daytona says sandboxes start in under 90 ms, with snapshots, persistence and GPU sandboxes on NVIDIA and AMD.
No plan fee. vCPU $0.0504 an hour, memory $0.0162 per GiB-hour, storage $0.000108 per GiB-hour after the first 5 GiB; $200 of free compute.
Sources:daytona.iodaytona.iodaytona.io
Firecracker microVM sandboxes on Vercel, billed by active CPU rather than wall time.
- Best for
- Agents whose code mostly waits on models or the network, and teams already on Vercel.
- Falls short
- Hobby is for personal, non-commercial use, and memory bills for idle time.
How it works and full pricing
CPU time spent waiting on I/O, such as a model’s reply, isn’t billed; memory is billed for the time it’s provisioned. Sessions run up to 24 hours on Pro, and persistent sandboxes resume across sessions.
Hobby includes 5 active CPU hours, 420 GB-hours of memory and 5,000 creations a month, 45-minute sessions. Pro ($20 a month with a $20 credit): $0.128 per active CPU hour, $0.0212 per GB-hour, $0.60 per million creations.
Sources:vercel.comvercel.com
Cloudflare Sandboxesby Cloudflare
From $5 a month plus usageContainers started and driven by a Durable Object, from Worker code.
- Best for
- Agents already built on Cloudflare Workers or the Agents SDK.
- Falls short
- Three meters (Containers, Workers, Durable Objects) to add up, and instance types top out at 4 vCPU and 12 GiB.
How it works and full pricing
A Durable Object starts a container and runs commands in it, with file transfer, preview URLs for servers it runs, R2 mounts and snapshots.
Workers Paid ($5 a month), billed as Containers: 375 vCPU-minutes, 25 GiB-hours and 200 GB-hours of disk included, then $0.000020 per vCPU-second, $0.0000025 per GiB-second and $0.00000007 per GB-second of disk, plus Workers and Durable Objects usage.
Sources:developers.cloudflare.comdevelopers.cloudflare.comdevelopers.cloudflare.com
Cloud browsers
Gives the agent a real browser in the cloud for sites with no API, with logins, proxies and CAPTCHAs handled.
NoClick’s browsers keep a person’s sign-ins and hand a CAPTCHA or a two-factor code to a person with a live view. Browsers in the docs
Hosted headless browsers with Stagehand, its SDK for AI browser agents.
- Best for
- Most teams starting with browser agents, especially with Stagehand.
- Falls short
- The free plan is one browser hour, and proxy traffic is $10 to $12 a GB.
How it works and full pricing
Your code or agent drives isolated browser sessions, with Stagehand’s act, extract and observe, live view, session replays and website sign-in.
Free: 1 browser hour, 3 at once. Developer: $20 a month, 100 hours then $0.12 an hour, 25 at once, 1 GB proxy then $12/GB. Startup: $99 a month, 500 hours then $0.10 an hour, 100 at once.
An open-source browser API for AI agents, also run as a hosted service.
- Best for
- Teams that want an open-source browser API they can also run themselves.
- Falls short
- Sessions are capped at 15 minutes until the $250-a-month Scale plan.
How it works and full pricing
Puppeteer or Playwright connect to its browsers over WebSocket, with stealth, residential proxies, CAPTCHA solving and persistent profiles, from REST, a CLI, SDKs or an MCP server.
Launch: $0 plus usage, $30 one-time credit, $0.10 per browser hour, 10 at once, 15-minute sessions, proxies $10/GB, CAPTCHAs $3 per 1,000. Scale: $250 a month with $100 credit, $0.08 an hour, 100 at once, 1-hour sessions.
Sources:steel.devdocs.steel.dev
Browser Use Cloudby Browser Use
$0.02 per browser hourHosted browsers plus Browser Use’s own browser agent and models.
- Best for
- Running many browser tasks at the lowest listed hourly rate.
- Falls short
- Concurrency grows only with what you’ve paid in total, and its agent pricing assumes its own models or a 0.2x fee on your key.
How it works and full pricing
You run its agent on a task, or drive the browser yourself; CAPTCHA solving is included and pay-as-you-go needs no subscription.
No subscription. Browsers $0.02 an hour, billed by the minute; home-IP proxies $5/GB; agent tasks from $0.006 (bu-ultrafast, median). 10 at once at first, up to 1,000 as total spend grows; $1 signup credit.
Sources:browser-use.com
Anchor Browserby Anchor
Free tier; $50 a monthCloud browsers for agents with authenticated sessions and anti-bot features.
- Best for
- Logged-in work on sites with strong bot detection.
- Falls short
- Plans are priced in credits, and browser-hour and proxy rates appear only in calculator examples, so a month is hard to predict.
How it works and full pricing
Paid plans add automated CAPTCHA bypass, authenticated browsers and any geolocation; Team adds Cloudflare Verified Browser Agents and custom proxies.
Free: 5 credits a month, 5 at once. Starter: $50 a month, 50 credits, 25 at once; overage $1 per credit; each completed task 0.1 credits. Its calculator examples show $0.09 per browser hour and $8 per proxy GB.
Sources:anchorbrowser.io
Cloud browsers billed by the GB-second, headless or headful, with no charge for idle time.
- Best for
- Cheap headless browsing, and headful browsers when sites need them.
- Falls short
- Headful browsers, which bot-checking sites need, cost 8 times headless, and regional browsers need the $200-a-month plan.
How it works and full pricing
Free plans include stealth mode with a CAPTCHA solver and a managed residential proxy; Hobbyist adds unlimited managed auth and vaults; Start-up adds regional browsers and your own proxies.
Headless $0.0000166667 a second (about $0.06 an hour); headful $0.0001333336 a second (about $0.48 an hour). Free: $0 plus usage, $5 a month credit, 5 at once. Hobbyist: $30 a month, $10 credit, 10 at once. Start-up: $200 a month, 150 at once.
Email inboxes
An address the agent owns: it receives mail (customer email, sign-up codes), replies on the thread, and gets an event when mail arrives.
NoClick gives an agent its own inbox in one call. Email for AI agents
An email API built for agents: inboxes made by API on @agentmail.to or your own domain.
- Best for
- Products where many agents each need a full inbox of their own.
- Falls short
- Email only (no calendar or phone), and past the Developer plan the next tier is $200 a month.
How it works and full pricing
An inbox sends, receives, replies, forwards and threads mail, with events by webhook or WebSocket (no public URL needed), plus IMAP and SMTP, SDKs and an MCP server. Reply extraction returns only the new text of each message.
Free: 3 inboxes, 3,000 emails a month. Developer: $20/month for 10 inboxes and 10,000 emails, then $2/month per inbox and $2 per 1,000 emails. Startup: $200/month for 150 inboxes.
Sources:agentmail.todocs.agentmail.to
Email and calendar identities for agents, on the same API Nylas uses for users’ own Gmail and Outlook accounts.
- Best for
- Agents that need a calendar as well as an inbox, or products that also read users’ own mail.
- Falls short
- Production mail runs on a domain you control, so DNS setup comes first.
How it works and full pricing
Each Agent Account is a real mailbox with system folders plus a primary calendar that hosts events and answers invitations. It runs on your domain or a *.nylas.email trial domain, and new mail fires a message.created webhook.
Free: 3 Agent Accounts, 3,000 emails a month. Essentials: $15/month for 10 accounts and 10,000 emails. Pro: $49/month for 50 accounts, then $0.40 an account.
Sources:nylas.comdeveloper.nylas.com
Resend Inboxesby Resend
3 free; $20 a monthAn email sending API that also hosts inboxes on your verified domains.
- Best for
- Teams already sending app email through Resend.
- Falls short
- Inboxes need your own domain with receiving enabled: there’s no shared domain to start on.
How it works and full pricing
An inbox keeps the threads of one address in inbox, sent, archive, trash and spam folders; you send from it and answer inbound mail in the same thread. Separately, inbound email can go to your own webhook with content and attachments parsed.
Free: 3,000 emails a month (100 a day) and 3 inboxes. Pro: $20/month for 50,000 emails; extra inboxes $1/month each. Received email counts toward the quota.
Sources:resend.comresend.com
Postmark Inboundby Postmark
$16.50 a monthTransactional email with inbound processing that turns each received email into JSON for your app.
- Best for
- Reliable sending plus a simple inbound webhook you build an agent inbox on.
- Falls short
- No inboxes or threads: storing and threading messages per agent is your code.
How it works and full pricing
Use a generated inbound address or your own domain, and Postmark posts each message to your webhook as JSON with headers, a clean text version of the reply, spam score and attachments. Your code stores, threads and answers it.
Pro: $16.50/month from 10,000 emails, inbound processing included (the $15 Basic plan doesn’t list it). Free developer tier: 100 emails a month.
Sources:postmarkapp.compostmarkapp.com
Phone numbers, SMS and voice
A real number the agent answers and calls from, texts on, and a voice layer that turns a caller’s speech into turns of your agent.
NoClick buys a number and has your agent answer it, the voice included in the minute. Phone numbers for AI agents
The largest communications API: numbers, SMS, voice and WhatsApp, plus ConversationRelay for voice AI.
- Best for
- Teams that want voice, SMS and WhatsApp under one account with the widest reach.
- Falls short
- You build and host the agent behind the WebSocket, and US texting needs A2P 10DLC registration, whose fees the pricing page doesn’t list.
How it works and full pricing
Buy a number by API; ConversationRelay does speech-to-text and text-to-speech and sends the caller’s words to your app over a WebSocket you host, where your AI writes the reply.
US local number $1.15/month. Calls $0.0085/min in, $0.0140/min out. SMS $0.0083 a segment each way plus carrier fees. ConversationRelay $0.07/min.
Sources:twilio.comtwilio.comtwilio.com
A carrier on its own network: numbers, SMS, SIP and a voice AI agent product.
- Best for
- Lower carrier prices with a voice agent on the same network.
- Falls short
- Each call rounds up to the next full minute, so short calls cost more than the per-minute rate suggests.
How it works and full pricing
Buy numbers and route calls by API or SIP. Its voice AI engine runs the conversation (turn-taking, tools, speech-to-text, text-to-speech), with the model billed per token and the call at carrier rates.
US local number from $1/month (SMS adds $0.10/month). SMS $0.004 a part plus carrier fees. Voice AI engine $0.05/min plus model tokens and SIP minutes (from $0.0032/min in).
Sources:telnyx.comtelnyx.comtelnyx.com
A communications API for SMS and voice at lower per-message rates than Twilio.
- Best for
- Cost-sensitive SMS at volume.
- Falls short
- A carrier API, not an agent layer: the voice AI and the agent’s logic are yours to build.
How it works and full pricing
Rent a long code and send and receive SMS through its API.
US long code $0.50/month. SMS $0.0077 a message each way, plus carrier surcharges.
Sources:plivo.com
A developer platform for voice agents assembled from a transcriber, a model and a voice you choose.
- Best for
- Teams that want to choose and tune each voice component.
- Falls short
- Concurrency is capped by plan (4 calls on pay-as-you-go), and each extra line is $10 a month.
How it works and full pricing
You pick the speech-to-text, LLM and voice providers; Vapi hosts the call on its own telephony or on a Twilio, Vonage or Telnyx number, and passes model costs through at cost.
$0.05/min hosting plus models at cost: $82 to $129 for 1,000 minutes in its calculator. $5 free credit and 1 number; Core plan $29/month for 5 numbers and 10 concurrent calls.
Sources:vapi.ai
A hosted voice agent platform with templates, call analytics and simulation testing.
- Best for
- Teams that want a voice agent with testing and analytics ready-made.
- Falls short
- The per-minute price spans more than four times from low to high, so price your own model and voice before you commit.
How it works and full pricing
Build the agent in its dashboard or API, give it a Retell number, and it answers and places calls with transcripts and analytics included.
$0.07 to $0.31 a minute for voice agents. Numbers $2/month. $10 free credit; 20 concurrent calls free, then $8 each a month.
Sources:retellai.com
Connected accounts
Each of your users connects their own Gmail, Slack or calendar once, and the agent acts in exactly that account while tokens are stored and refreshed for you.
NoClick sends each user a connect page and keeps their tokens fresh. Connections in the docs
A tool platform for agents with managed auth across 1,500+ toolkits.
- Best for
- Giving an agent many apps as ready-made tools quickly.
- Falls short
- In production you register your own OAuth app with each provider, so each provider’s app review is still yours.
How it works and full pricing
Each of your user ids holds connected accounts made through a hosted Connect Link; tools run in that account, and Composio refreshes OAuth tokens before they expire. Its own OAuth apps are for development; its docs say to use your own OAuth credentials in production.
Hobby: free, 100,000 tool calls a month. Pro: $29/month as credits, then $0.0003 a call; connections free on your own OAuth apps, 1,000 free on Composio’s, then $0.10 each once.
Sources:composio.devdocs.composio.dev
Integration infrastructure: managed auth for 1,200+ APIs plus pre-built tools, triggers and syncs.
- Best for
- Engineering teams that want code-level control over each integration.
- Falls short
- Per-connection pricing grows with your users: 1,000 users with two apps each is about $580 a month in connections.
How it works and full pricing
Your users connect their accounts, Nango keeps the credentials, and your agent uses its 10,000+ pre-built tools, triggers and syncs or your own integration code run on Nango’s compute.
Free: 10 connections and 10 compute hours a month, hard-capped. Pay-as-you-go: $50/month including $50 of credits; $0.29 per connection a month, $0.72 per compute hour.
Sources:nango.dev
An actions runtime that runs agent tools as the signed-in user, with 8,000+ permission-aware tools.
- Best for
- Agents where every action must run with the user’s own permissions.
- Falls short
- At $0.01 a call, an agent making 100,000 calls a month pays $1,000 in call fees.
How it works and full pricing
The agent calls Arcade’s tools, which act on behalf of the authenticated user, “never through broad service accounts”, and enforce that user’s permissions on each action.
Free: 2,000 tool calls a month. Team: $25/month plus $0.01 a tool call.
Sources:arcade.devarcade.dev
Pipedream Connectby Pipedream
$150 a monthPipedream’s way for your users to connect accounts in your app, with 3,000+ connectors and 10,000+ tools and triggers.
- Best for
- The widest connector catalog, priced per user rather than per connected account.
- Falls short
- No free production tier: $150 a month before your first user is live.
How it works and full pricing
Users connect accounts through Connect; your app or agent runs actions and MCP tools in those accounts or calls the API through the Connect proxy, each using credits for its compute time.
Free in development only. Startup: $150/month billed monthly ($99 billed annually) with 100 external users, then $2 a user a month; 10,000 credits, then $0.012 each (a credit is 30 seconds of compute).
Sources:pipedream.com
Auth0 Token Vaultby Auth0 (Okta)
In the Free planAuth0’s store for users’ Google, Microsoft, Slack, GitHub and other provider tokens, part of Auth0 for AI Agents.
- Best for
- Apps already on Auth0 that want provider tokens tied to their users’ login.
- Falls short
- It holds tokens, not tools: you write each API call yourself.
How it works and full pricing
Users connect accounts through Auth0, which keeps the access and refresh tokens; your agent exchanges the user’s Auth0 token for that provider’s token and calls the provider’s API itself.
Token Vault is on the Free plan (2 included, up to 25,000 MAUs). The Auth0 for AI Agents add-on, with unlimited Token Vault, adds 50% to the base: B2C Essentials is $35/month plus $18 at 500 MAUs.
MCP server hosting
Puts an agent’s tools at one URL that Claude, ChatGPT, Cursor or your own agent can call over MCP.
NoClick hosts an MCP server over your users’ connected apps, web search and page fetch, free. Remote MCP servers
Cloudflare Workers (remote MCP)by Cloudflare
Free tier; $5 a monthYour MCP server as a Worker, over Streamable HTTP.
- Best for
- A public MCP server with OAuth for outside users, at the lowest cost.
- Falls short
- You write for the Workers runtime, and McpAgent is deprecated, so older servers need migrating.
How it works and full pricing
Templates use createMcpHandler, which builds a new server per request, so no session state is needed; state that must last goes in KV, D1 or a Durable Object. The Workers OAuth Provider library signs users in and validates tokens.
Workers Free: 100,000 requests a day. Workers Paid: $5 a month minimum with 10 million requests and 30 million CPU ms, then $0.30 per million requests and $0.02 per million CPU ms.
MCP servers as Vercel Functions on Fluid compute, built with the mcp-handler package.
- Best for
- Teams already shipping a Next.js app on Vercel.
- Falls short
- mcp-handler doesn’t issue tokens or run an authorization server, and Hobby is personal, non-commercial use only.
How it works and full pricing
A route in a Next.js app serves MCP over Streamable HTTP; withMcpAuth verifies bearer tokens and publishes OAuth protected-resource metadata, with previews and instant rollback.
Hobby: $0 with 4 active CPU hours, 360 GB-hours and 1 million invocations a month. Pro: $20 a month per developer seat with a $20 credit, then from $0.128 per active CPU hour, $0.0106 per GB-hour and $0.60 per million invocations.
Sources:vercel.comvercel.com
Google Cloud Runby Google Cloud
Free tier, then usageYour MCP server as a container, scaling with requests.
- Best for
- Internal MCP servers for teams on Google Cloud, behind IAM.
- Falls short
- IAM is the default, so OAuth for outside clients is yours to build.
How it works and full pricing
Cloud Run serves MCP over Streamable HTTP or SSE from a container image or source; requests need IAM authorization by default.
Free each month: 180,000 vCPU-seconds, 360,000 GiB-seconds and 2 million requests. Then $0.000024 per vCPU-second, $0.0000025 per GiB-second and $0.40 per million requests (us-central1, request-based billing).
Hosting and a gateway for MCP servers from the makers of FastMCP; fastmcp.cloud now redirects here.
- Best for
- Python servers built with FastMCP that need OAuth without running an auth server.
- Falls short
- Usage rates for the Developer plan aren’t on the pricing page.
How it works and full pricing
Remote MCP hosting with Git-based CI/CD and managed OAuth, with connections to Claude, Claude Code and Cursor; paid plans add analytics and your own auth.
Personal: free, 1 developer, 200 MCP servers, 1 agent, 1-day logs. Developer: $35 per developer a month plus usage.
Sources:horizon.prefect.ioprefect.io
Your MCP server as a Machine from a Docker image.
- Best for
- Any language or a stdio server wrapped for HTTP, on a small always-cheap VM.
- Falls short
- Auth, scaling and the container are yours, and a suspended Machine wakes on the first call.
How it works and full pricing
fly launch writes the config; the server answers Streamable HTTP at /mcp, and auto-stop suspends the Machine when idle and wakes it on the next request. Auth is OAuth 2.1 or a bearer token you add.
A 256 MB shared-cpu-1x Machine is $2.19 a month running all day; Fly.io puts it at about $0.28 to $0.83 a month at 2 to 8 hours a day. No plan fee.
Spending and payments
Lets an agent pay for things within limits you set, with a person approving what’s above them.
Spending limits get a guide of their own. Spending limits for AI agents
Virtual cards issued per agent or per task, with Stripe’s spend controls.
- Best for
- Businesses that want card-level control and write their own approval logic.
- Falls short
- You apply first, then build the authorization webhook and any human review yourself.
How it works and full pricing
Each agent gets its own card with limits per purchase or period and allowed merchant categories; every purchase sends a real-time authorization webhook you approve or decline within 2 seconds. Single-use cards and Shared Payment Tokens cover browser and programmatic checkout.
$0.10 per virtual card in the US. Issuing needs an application; other Issuing fees not published on the pages we could load.
Sources:docs.stripe.comdocs.stripe.com
Ramp’s spend management extended to AI agents.
- Best for
- Companies already running their spend on Ramp.
- Falls short
- Beta access through Ramp’s sales team, with no published price.
How it works and full pricing
Agents get auto-expiring cards and a key tied to a person on your team; limits are set per agent, task or merchant, routine spend is auto-approved and anything important is held for a person.
Not published for Ramp for Agents (beta access through Ramp’s team). Ramp itself: Free plan $0/user/month; Plus $15/user/month plus a platform fee.
Sources:agents.ramp.comramp.com
x402by Coinbase
1,000 free a month, then $0.001 eachAn open protocol for paying per request over HTTP 402, with Coinbase’s facilitator settling payments onchain.
- Best for
- Agents paying small amounts per call to APIs that accept x402.
- Falls short
- Works only with sellers that accept x402, and the agent needs a funded crypto wallet.
How it works and full pricing
A paid API answers that payment is required, with the price and payee; the agent signs a payment and retries with proof attached. USDC is the most common asset, on Base, Solana and other networks.
Coinbase’s facilitator: the first 1,000 onchain transactions a month are free, then $0.001 each; verifying a payment is free.
Sources:docs.cdp.coinbase.comdocs.cdp.coinbase.comdocs.cdp.coinbase.com
Start with a key.
Create an API keyOr paste one prompt into your coding agent. It saves a key to your .env.
Get the prompt in the consoleOr add NoClick’s MCP server to your client.
Questions
If it should sleep between messages, Cloudflare Agents gives each agent a Durable Object with its own SQLite storage that isn’t billed for duration while hibernating, from $5 a month. If it’s a process you wrote, a small always-on VM works: a 256 MB Fly.io Machine is $2.19 a month, or Railway from $5 with Postgres beside it. On AWS, Bedrock AgentCore runs each session in its own microVM for up to 8 hours, with memory as a separate service.
By listed rate, boat is lowest: $0.036 an hour for a VM with 4 vCPU and 8 GB, billed per second, with a $20 monthly minimum. E2B and Daytona both charge $0.0504 per vCPU-hour and $0.0162 per GiB-hour, with $100 and $200 of free credit and no minimum. Vercel Sandbox bills only active CPU ($0.128 an hour), which can cost less when the code mostly waits on a model. Modal costs more per core but adds GPUs and 24-hour sandboxes.
Browserbase is the common default: $20 a month for 100 browser hours, with Stagehand for act and extract. Steel is open source at $0.10 an hour, Browser Use Cloud lists the lowest rate at $0.02 an hour, and Kernel runs headless browsers for about $0.06 an hour. For sites with strong bot checks, compare their stealth, CAPTCHA and proxy prices, which vary more than the hourly rates.
Serve it over Streamable HTTP from a serverless host: Cloudflare Workers (createMcpHandler and an OAuth provider library, 100,000 free requests a day), Vercel with mcp-handler, or Google Cloud Run behind IAM; Prefect Horizon hosts FastMCP servers with managed OAuth on a free personal plan. If the tools are your users’ connected apps rather than your own code, NoClick hosts that server for free. Smithery’s hosted server URLs stopped working on October 31, 2026.
NoClick’s inboxes are free to create and $0.00035 a reply, with the agent woken by each email. Among the specialists: AgentMail if many agents each need a full inbox ($20/month for 10 on Developer), Nylas Agent Accounts if they also need a calendar ($15/month for 10), Resend Inboxes if you already send through Resend, and Postmark if you only need inbound mail delivered as JSON.
On NoClick a US number is $1 a month and $0.0607 a minute answered, voice included, plus the agent’s model calls. Or buy a number from a carrier like Twilio or Telnyx and connect your agent through a voice layer such as ConversationRelay ($0.07/min) or Telnyx’s voice AI ($0.05/min), or use a hosted voice platform like Vapi or Retell AI at about $0.07 to $0.31 a minute.
For business uses like support, sales and notifications, yes, through Twilio, Telnyx or another provider of Meta’s API. Meta’s WhatsApp Business Platform terms (clause 4.7, AI Providers) bar providers of general-purpose AI assistants from the platform, and the rule is Meta’s, so it holds through any provider.
Use a platform that runs per-user OAuth and stores and refreshes the tokens: Composio and Arcade charge per tool call, Nango per connection, Pipedream Connect per user, and Auth0 Token Vault per monthly active user. In production you usually register your own OAuth app with each provider, so their app reviews are still yours.