Back to vendors
n

ngrok

Also known as: ngrok.com, ngrok AI Gateway, ngrok.ai, ngrok MCP Server

Visit site
Entry priceAI Gateway $0.05 per million tokens routed; Free with a $5 credit; Hobbyist $10 a month; Pay as you go $20 a month plus usage; Enterprise quotedFull pricing detail

ngrok is developer infrastructure that routes and secures traffic to apps, APIs and AI models wherever they run. Its AI Gateway sits between an agent and model providers such as OpenAI, Anthropic, Google, xAI and DeepSeek, with virtual models, fallbacks, key level failover, per key model access and an AI Firewall that blocks risky agent actions. The same network exposes local MCP servers behind token, OAuth or JWT checks and connects agents to data inside customer networks.

ngrok started as the tunnel developers use to put a local app on the internet and has grown into a gateway platform: secure tunnels and cloud endpoints, an API gateway with a declarative Traffic Policy layer for authentication, rate limiting, a WAF and routing, private cross network connectivity, and products for AI traffic. Developers and companies use it to receive webhooks, expose services, and connect software running in one network to systems in another, and teams at GitHub, Okta, Databricks, Twilio, Zoom and OpenAI are among its users.

For agent builders, the AI Gateway is a managed proxy at gateway.ngrok.ai that works with the OpenAI, Anthropic and Vercel AI SDKs by changing one URL. It routes to OpenAI, Anthropic, Google, xAI, DeepSeek, Groq, Mistral, Moonshot, Qwen, Tenstorrent, Ollama, custom providers and local models over private connectivity, with virtual models that point one name at an ordered list of models, automatic retries, rerouting when a model or key slows or fails, and spare keys that take over at rate limits. Provider keys stay on the gateway, separate access keys go to each app or developer with rules about which providers and models each may call, and usage rolls up into tokens, latency and errors by app, developer and model.

An AI Firewall inspects agent actions at the gateway before they run, catching prompt injection phrases, exposed credentials, destructive shell commands and force pushes. ngrok also puts local MCP servers on the internet for Claude, ChatGPT, Cursor and other clients, with bearer tokens, secret paths, IP categories, OAuth pass through or JWT validation enforced at the edge, and Agent Connect links agents to data inside customer networks. An ngrok MCP Server lets Claude Code, Codex, Cursor and others inspect traffic, validate policies and manage the account.

The AI Gateway charges $0.05 per million tokens routed, with inference billed by each provider on the customer's own keys. Core plans run from Free with a $5 credit to Hobbyist at $10 a month and Pay as you go at $20 a month plus usage, with Enterprise quoted. Traffic runs through eight points of presence, from Ohio and California to Frankfurt, Mumbai, Tokyo and São Paulo, and endpoints can be pinned to chosen regions. ngrok holds a SOC 2 Type II report, offers HIPAA with a BAA, and sells RBAC, SSO, SCIM, audit logs and region specific routing as add ons or Enterprise features.

It fits developers and platform teams that want one network layer for their agents' model calls, MCP servers and private data, with failover and access rules outside the application code. It fits less well for teams looking for an agent builder, a workflow engine or memory and retrieval, none of which ngrok provides.

Vendor details

Canonical URL

https://ngrok.com

Category

Agent infrastructure

Subcategory

Ingress, API gateway and AI gateway that route agent traffic to models, MCP servers and private networks under traffic policy

Funding status

Private company, headquartered in San Francisco and founded by Alan Shreve, who released the first version in 2013. Its first outside round, a $50 million Series A led by Lightspeed Venture Partners with Coatue, closed in December 2022.

Company status

independent

Use cases & customers

Primary use cases

routing agent model calls across providers with fallbacksper key access rules and usage tracking for LLM trafficexposing local MCP servers to AI clients behind authenticationconnecting agents to data inside customer networksblocking risky agent actions with an AI Firewall

Target customers

developers building AI agents and appsplatform and infrastructure teamscompanies connecting agents to customer networksteams exposing MCP servers

Deployment options

Managed cloud gateway across eight points of presenceAgent and SDKs running in the customer's networkRegion pinning and region specific routing

Integrations

The AI Gateway works with the OpenAI, Anthropic and Vercel AI SDKs and routes to OpenAI, Anthropic, Google, xAI, DeepSeek, Groq, Mistral, Moonshot, Qwen, Tenstorrent, Ollama and custom providers. MCP exposure supports Claude Code, Cursor, ChatGPT and Claude.ai connectors, and the ngrok MCP Server connects Claude Code, Codex, Cursor, Gemini CLI, Amp, Cline and VS Code to the account. The ngrok API manages endpoints, domains, keys and policies.

In practice

A startup points its OpenAI SDK at the ngrok AI Gateway and defines a virtual model that tries its self hosted Llama server first, then Claude, then GPT. When the local GPU box goes down overnight, requests reroute without a deploy.

A platform team gives each internal app its own gateway key limited to two approved providers, and finance sees token spend broken down by app and developer instead of one provider bill.

A developer exposes a local MCP server to Claude.ai through ngrok with a JWT validation rule from the company's identity provider, so only signed in employees reach the tools.

Agentic Index coverage score

7.0 / 14 capabilities · 50%

Integrations & Tool Calling Full

ngrok carries an agent's authenticated calls to the systems it acts in. Local MCP servers go online for Claude Code, Cursor, ChatGPT and Claude.ai connectors with bearer tokens, secret paths, IP categories, OAuth pass through or JWT validation enforced at the edge, Agent Connect links agents to data inside customer networks without inbound ports, and internal endpoints keep private services reachable only through a policy controlled cloud endpoint.

Sourcengrok.com/docs/gateway/api-gateway/mcp-gatewayread 2026-11-18

Workflow Orchestration Not documented

Traffic Policy chains actions on each request and response, such as deny, rate limit, URL rewrite, JWT validation and forwarding to an internal endpoint, and the AI Gateway retries and falls back across models. Those are per request rules in the network path. Planning steps, running multistep work or coordinating several agents stays in the customer's own agent code.

Sourcengrok.comread 2026-11-18

Knowledge Grounding & RAG Not documented

Retrieval sits outside ngrok's scope. The gateway forwards prompts and responses between the customer's app and the model provider, and Agent Connect opens a path to where the customer's data lives, but ngrok keeps no index, vector store or knowledge base of its own for an agent to search. Any grounding happens in the customer's app before a prompt reaches the gateway.

Sourcengrok.airead 2026-11-18

Human Oversight & Guardrails Partial

An AI Firewall inspects agent actions at the gateway before they run and stops ones that match risky patterns, such as prompt injection phrases, exposed credentials, destructive shell commands and force pushes. Traffic Policy can deny or answer requests at the edge so they never reach a private MCP server. Blocking is automatic; nothing pauses an action for a person to approve.

Sourcengrok.airead 2026-11-18

Security, Identity & Governance Full

ngrok holds a SOC 2 Type II report, with a DPA, subprocessor list and penetration test summary in its trust center, and signs HIPAA business associate agreements. Teams add SSO and role based access control, SCIM provisioning and audit logs, and the AI Gateway keeps provider keys on the gateway while issuing separate access keys per app or developer with rules on which providers and models each may call.

Sourcengrok.com/pricingread 2026-11-18

Observability & Auditability Full

Every call routed through the AI Gateway rolls up into tokens, latency and errors attributed to the app, developer and model that made it. The traffic inspector shows each request and response, including the JSON RPC bodies of MCP calls, traffic logs export at $0.25 per 2,000 events, and audit logs record account changes. A costly model or a noisy caller shows up in that breakdown.

Sourcengrok.airead 2026-11-18

Memory & State Persistence Not documented

Each request passes through the gateway on its own, with usage counted for reporting. Conversation state and anything an agent should remember between sessions stay in the customer's application, since ngrok stores no memory, session history or user context for agents. Each new request carries only what the customer's own code chooses to send along with it.

Sourcengrok.airead 2026-11-18

Deployment & Data Residency Full

Traffic runs through eight points of presence in Singapore, Sydney, Frankfurt, Mumbai, Tokyo, São Paulo, Ohio and California. Customers can pin an endpoint to chosen regions through the agent's region setting, a domain's resolves_to property or global load balancer configuration, and region specific routing restricts a domain to chosen geographic regions as an add on billed per active hour. The ngrok agent and SDKs run inside the customer's own network, and the AI Gateway can route to models the customer hosts itself over private connectivity.

Sourcengrok.com/docs/universal-gateway/points-of-presenceread 2026-11-18

Prebuilt Agents / Templates / Packs Not documented

Guides show example Traffic Policy rules for authentication, rate limiting, WAF and MCP exposure that teams copy into their own configuration, and the gateway ships with named providers ready to route. There are no prebuilt agents, agent templates or packaged playbooks, since ngrok supplies the network around agents rather than agents themselves. Each agent is the customer's own.

Sourcengrok.com/docs/gateway/api-gateway/mcp-gatewayread 2026-11-18

Triggers & Channel Coverage Partial

Requests and webhooks reach a customer's agent through ngrok endpoints, with stable domains that survive restarts even on the free plan and Traffic Policy deciding what gets through. ngrok delivers the inbound event but does not start anything itself, since it has no schedules, queues or event subscriptions that wake an agent on their own. Waking stays with the customer's code.

Sourcengrok.comread 2026-11-18

Model Flexibility & Routing Full

Developers name the model in each request and choose among OpenAI, Anthropic, Google, xAI, DeepSeek, Groq, Mistral, Moonshot, Qwen, Tenstorrent, Ollama, custom providers and local models. Virtual models point one name at an ordered list so teams change models without a deploy, fallbacks and key level failover take over when a model or key fails, access rules limit models per key, and customers bring their own provider keys.

Sourcengrok.airead 2026-11-18

APIs / SDKs / MCP Extensibility Full

The ngrok API manages the account programmatically, and the ngrok MCP Server, reached over OAuth, lets Claude Code, Codex, Cursor, Gemini CLI, Amp, Cline and VS Code check usage, query the traffic inspector, validate traffic policies and read or write endpoints, domains, API keys, IP policies, credentials and vaults. The AI Gateway itself takes the OpenAI, Anthropic and Vercel AI SDKs unchanged apart from the base URL.

Sourcengrok.com/mcpread 2026-11-18

Testing, Debugging & Optimization Not documented

Teams debug traffic by replaying and inspecting requests and can validate a Traffic Policy before it goes live through the MCP server, which checks that configuration is well formed. Nothing in ngrok scores model or agent output, runs test sets or compares versions of an agent's behavior. The quality of an agent's answers is measured outside ngrok, in the customer's own tooling.

Sourcengrok.com/mcpread 2026-11-18

Browser / Computer-use Not documented

Moving network traffic is the whole job, so there is no browser, desktop session or remote computer for an agent to operate. An agent that needs to click through a website does that with its own tools, and ngrok only provides the connection path. The same holds for desktop and terminal work: ngrok carries the traffic and leaves every action to the agent at either end.

Sourcengrok.comread 2026-11-18

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

AI Gateway $0.05 per million tokens routed; Free with a $5 credit; Hobbyist $10 a month; Pay as you go $20 a month plus usage; Enterprise quoted

A monthly plan fee plus metered endpoint hours, data transfer, requests, connections and policy units, with AI Gateway tokens billed per million routed

Free tier

Included quota

Hobbyist and Pay as you go each include 5 GB of transfer and 100,000 HTTP requests, with $10 and $20 of monthly credit. Pay as you go includes 3 users, then $5 per user.

What is public

ngrok's AI Gateway charges $0.05 per million tokens routed on every model and provider, and providers bill inference directly. Platform plans run from Free with a $5 credit to Hobbyist at $10 a month and Pay as you go at $20 a month plus usage. Enterprise is quoted.

Billing mechanics

Each plan includes a monthly usage credit, and metered charges apply beyond it. Identity and governance features are priced per user and apply to every user on the account.

Cost watchouts

Model inference is billed separately by each provider on the customer's own keys. SSO and RBAC add $10 per user a month for every user, and the governance suite with SCIM and audit logs another $15. Dedicated Agent Connect IPs cost $900 a month per region.

Variable cost rationale

Gateway fees scale with tokens routed and traffic, though the per token rate is small next to provider inference.

Additional watchouts

Region specific routing is listed at $0.02 per hour with billing on Pay as you go not yet switched on.

Overage / add-ons

Usage beyond the credit bills at $0.02 per endpoint hour, $0.10 per GB out, $1 per 100,000 HTTP requests, $2 per 100,000 TCP connections and $0.10 per 100,000 Traffic Policy units.

Sales call required

Mixed (some tiers require a call)

Free / trial

Free plan with a one time $5 usage credit

Lowest paid plan

Hobbyist, $10 a month with a $10 monthly usage credit

Commercial notes

At the AI Gateway rate, routing 250 million tokens costs $12.50 in gateway fees.

Missing data

ngrok gives no Enterprise price, and the per hour rate for region specific routing does not yet bill on Pay as you go.

Agentic Index verified 2026-11-18

Alternatives to ngrok

Other products that do the work ngrok does, and what each one is.

Contact us

Found a vendor we missed? Have feedback on the index? We'd love to hear from you.