Best AI Gateways With Managed Agent Tools 2026

Amazon Bedrock AgentCore Gateway is the strongest managed-tool gateway overall; Vercel AI Gateway is the fastest way to give an agent production search. At one million monthly agent interactions, AgentCore's gateway layer can cost about $40.02 before model, web-search, and backend charges, while Vercel removes the separate Exa account and key but charges $7 per 1,000 searches after August 31, 2026.

The short answer: which gateway wins which job?

An AI gateway earns a place in an agent stack when it removes a control surface, not when it merely puts another endpoint in front of a model. For managed tools, the decisive questions are whether the gateway can turn existing services into tools, hold credentials, decide who may call them, help the agent find the right tool, and record what happened.

Pick Amazon Bedrock AgentCore Gateway when agents must call internal APIs and SaaS systems under one managed identity boundary. It can convert OpenAPI, Smithy, and Lambda sources into MCP-compatible tools, then broker both inbound and outbound authentication.

Pick Vercel AI Gateway when the immediate problem is web search. Its Exa integration removes a separate account, key, and billing relationship. That is a meaningful workflow reduction, but it is narrower than turning a company's internal APIs into governed tools.

Pick Portkey when the MCP servers already exist and the missing layer is authentication, approval, and audit. Pick TrueFoundry when deployment location, hosted stdio servers, virtual MCPs, or pre-call and post-call guardrails matter. Pick Kong when Konnect already controls the API estate. Pick Cloudflare when the team wants free gateway basics and is comfortable assembling managed tools on the separate Agents platform.

What managed agent tools means, and what it does not

A model gateway routes prompts to models. An MCP gateway controls how agents discover and call tools. Some products do both, but the two bills and the two failure modes remain separate.

This distinction matters because a model call usually reads information and produces an answer, while a tool call can change a system. A support agent may use a model to interpret a request, then use tools to read an account, issue a credit, and update a ticket. The model gateway can govern the first step without knowing whether the second step was authorized.

A credible managed-tool gateway handles several jobs:

• Provisioning: It makes an API, function, or remote MCP server available without every application team rebuilding the connection.
• Identity: It verifies the calling agent or user, then supplies the approved credential to the upstream tool.
• Policy: It limits which tool and action the caller may use, sometimes with a human approval step.
• Discovery: It helps the model select a relevant tool without loading every tool schema into the prompt.
• Audit: It records which identity invoked which tool with which parameters and result.

Managed does not always mean hosted execution. Portkey can proxy and govern a remote MCP server while another platform still runs it. Cloudflare Agents can host an agent and supply managed tools, while Cloudflare AI Gateway remains the model-control surface. Vercel can provision Exa Search natively, but that one integration does not give it AgentCore's OpenAPI-to-MCP transformation layer.

The practical consequence is a separate tool-plane bill. It contains gateway calls, search or browser usage, tool hosting, secrets work, logs, approvals, and the engineering owner who responds when a tool starts failing. A zero-markup model gateway can still leave most of that bill untouched.

This is also why the cheapest model route and the cheapest agent workflow are not always the same. The broader model-gateway cost comparison is useful when token routing is the decision. Managed agent tools require a second sheet.

How these gateways were picked

The ranking turns on four buyer questions. First, how much tool work the product manages: transformation, registry, credentials, discovery, execution, or A2A traffic. Second, whether access can be constrained to a person, agent, workspace, or tool. Third, whether a buyer can calculate the production bill from current public prices. Fourth, where the product's wall appears: a deployment boundary, a separate product surface, a request allowance, or a control-plane fee.

Pricing pages, product documentation, tier limits, and launch pages were live-verified on August 27, 2026. Each ranked section includes a current product or documentation screenshot. The products were not exercised with production traffic during this review, so no latency, uptime, tool-call success, or support result is presented as a test.

Six gateways made the cut because each is defensible for a different job. Products that offered only model routing, could not support a specific managed-tool judgment, or hid every useful buying number behind sales were cut. Public pricing did not guarantee a high rank, but it made the operating consequence easier to judge.

The order favors removal of engineering and security work over the longest feature list. A narrow integration can beat a broad platform when it removes the exact step a small team faces this week. A broad platform wins overall only when its extra surface replaces work the buyer would otherwise own.

1. Amazon Bedrock AgentCore Gateway: best overall for enterprise tool planes

Amazon Bedrock AgentCore Gateway is the best overall choice when an agent must reach a mixed estate of internal APIs, AWS functions, SaaS tools, other agents, and models through a managed control point.

Its strongest capability is transformation. AgentCore Gateway can take an OpenAPI definition, a Smithy model, or an AWS Lambda function and expose it as MCP-compatible tools. A platform team can put an existing leave-management API behind the gateway instead of asking every agent team to write and host a separate MCP adapter.

The identity layer is equally important. AgentCore handles inbound authentication for the caller and outbound authentication for the target. It supports OAuth flows, token refresh, and secure credential storage. The agent gets permission to call a tool; it does not need possession of the Salesforce, Slack, or Jira credential that makes the action possible. One-click integrations are available for Salesforce, Slack, Jira, Asana, and Zendesk.

As the catalog grows, semantic tool selection helps an agent find a suitable tool without carrying every complete schema in its context. That reduces prompt bulk and makes a catalog of many internal actions more usable. The gateway can also front HTTP services and A2A traffic, so the same policy boundary can expand beyond MCP.

The wall is operational gravity. AgentCore removes gateway infrastructure, but the buyer still has to understand AWS identity, regions, service permissions, logging, and the upstream systems being exposed. A team with no AWS operating model may save adapter code while creating a new cloud ownership problem. The product is most compelling when AWS is already an approved control plane or when managed identity and API conversion justify that move.

Best for: Enterprises turning existing APIs and functions into governed agent tools
Standout: OpenAPI, Smithy, and Lambda transformation plus inbound and outbound authentication
Pricing: $0.005 per 1,000 gateway API invocations, $0.025 per 1,000 Search API calls, and $0.02 per 100 indexed tools per month; no upfront commitment or minimum fee
Free trial: Usage-based; new AWS customers may receive up to $200 in AWS Free Tier credits

A sensible first AgentCore gateway deployment

The safest first deployment is one bounded action, not an entire API estate. Pick a workflow with a named owner, a reversible outcome, and a permission model the security team already understands.

1. 1

   Map one approved action

   Choose one OpenAPI service or Lambda function, such as reading a leave balance. Expose only the operation the first agent needs. A smaller initial tool surface makes authorization and failure review legible.

2. 2

   Define both identities

   Name the caller allowed to reach the gateway and the outbound credential allowed to reach the target. Treat these as separate decisions. The agent's identity should not become a copy of the upstream service credential.

3. 3

   Index descriptions, not the whole estate

   Add clear tool names and descriptions for the bounded set. Use semantic selection only after the descriptions make the intended choice unambiguous. A search layer cannot rescue two tools that appear to do the same thing.

4. 4

   Price and observe a small traffic slice

   Route a controlled share of the workflow, count Search API and InvokeTool calls, and keep model, web-search, runtime, and backend charges on separate lines. Expand only after the action success rate and authorization logs are acceptable.

The verdict flips away from AgentCore when the team already operates remote MCP servers and only needs a shared authentication and audit proxy. Portkey is simpler for that narrower boundary. It also flips when web search is the sole missing tool and deployment speed matters more than transforming internal systems. Vercel is the shorter path there.

2. Vercel AI Gateway: best for zero-setup managed search

Vercel AI Gateway is the fastest choice when an agent needs current web search and the costly step is creating another vendor account, key, integration, and billing path.

Exa Search is now a built-in AI Gateway tool. It works with any AI Gateway model and requires no separate Exa account or API key. The application passes gateway.tools.exaSearch() as a tool, the gateway sends the query to Exa, returns the results to the model, and keeps the tool loop moving until the model stops searching.

That changes the Monday task. A small team can add grounded search without opening another procurement path or wiring a second secret. The benefit is setup compression, not a permanent price discount. Vercel lists Exa Search at $7 per 1,000 queries and bills it at list price with no markup. The search is free only through August 31, 2026, so a production plan should already carry the post-promotion price.

At 100,000 searches per month, the Exa line becomes $700. Vercel charges no platform fee or markup on model tokens, but optional governance can add usage lines. Custom Reporting costs $0.075 per 1,000 writes and $5 per 1,000 reporting queries. A team-wide provider allowlist costs $0.10 per 1,000 successful requests on Pro and Enterprise, while the per-request only filter has no added cost. Team-wide zero-data retention also costs $0.10 per 1,000 requests on Pro and Enterprise, while per-request ZDR has no added cost on those plans.

The wall is breadth. Vercel has removed the provisioning step for a valuable search tool, but that does not make it a general transformation layer for a private OpenAPI estate. If the requirement says "turn our benefits, billing, and inventory APIs into discoverable tools with outbound OAuth," AgentCore or TrueFoundry is closer to the job.

Best for: Product teams adding managed search to an AI SDK application
Standout: Exa Search with no separate Exa account or key
Pricing: Free includes $5 monthly credit on eligible models; Paid uses purchased credits at provider list rates with zero token markup; Enterprise can use invoice billing and arranged volume discounts
Free trial: Free tier; buying credits ends the monthly $5 free credit

The Free tier includes $5 in monthly credit, a subset of eligible models, lower per-model limits, and no BYOK. Paid is pay as you go with purchased credits, opens the full model catalog, allows custom limits and BYOK, and has no commitment. Once credits are purchased, the monthly free credit stops. Enterprise teams can arrange invoice billing without a payment-processing fee and can discuss volume token-spend discounts.

Vercel earns second place because its new managed tool is immediately useful and unusually low-friction, not because it covers the most tool jobs. Choose it when search is the action that makes the agent useful. Skip it as the sole gateway when the agent must safely change internal systems.

3. Portkey: best for one authentication plane across remote MCP servers

Portkey is the best fit when MCP servers already exist and the missing product is a shared door for credentials, permissions, approvals, and audit.

The Portkey MCP Gateway sits between MCP clients and servers. A client authenticates once to Portkey. The gateway checks access to the requested server and tool, injects the approved upstream credential, proxies the call, and logs the request and response. Upstream authentication can use OAuth tokens, API keys, or identity headers, while callers can use Portkey API keys or tokens from an identity provider.

This is the clean answer for a platform team tired of distributing GitHub, Slack, or internal-service credentials to every agent workspace. The MCP Registry holds server configuration and authentication, then access can be provisioned to workspaces. Content filters and approval workflows create another control point before a sensitive action runs.

Portkey's boundary is clear: it manages access to MCP servers rather than serving as the managed runtime for every tool. The server still has to exist somewhere, remain available, and expose the action correctly. If the company's problem is converting an OpenAPI service or hosting a command-line MCP server, TrueFoundry or AgentCore removes more work.

The public entry price is attractive, but its unit is recorded logs. Developer is Free Forever and permits traffic after the log allowance, but logs beyond the allowance are not recorded. That makes the tier suitable for prototypes, not for proving who changed a customer record in production.

Best for: Platform teams centralizing access to existing remote MCP servers
Standout: One client authentication path with upstream credential injection and per-workspace provisioning
Pricing: Developer is free; Production is $49 per month; Enterprise is custom priced
Free trial: Free Forever Developer tier

Developer is Free Forever with 10,000 recorded logs per month, three days of log retention, and 30 days of metrics. Portkey explicitly labels it unsuitable for production. Production costs $49 per month for 100,000 recorded logs, 30 days of logs, 90 days of metrics, RBAC, service account keys, production support, and simple plus semantic caching. Each additional 100,000 requests costs $9. At one million recorded requests, that is $130 before model usage.

Enterprise uses custom pricing and starts at 10 million-plus recorded logs per month. It adds custom retention, SSO, granular budgets and rate limits, private cloud, VPC hosting, data-lake export, custom guardrail hooks, and advanced compliance controls.

The decision flips to Portkey when the tool estate is already remote and healthy. It flips away when hosting, transformation, or private deployment of the servers is the bigger workload. In that case, buying only the proxy leaves the hard task in the team's backlog.

4. TrueFoundry: best for private MCP control and hosted tool formats

TrueFoundry is the strongest choice for a buyer that needs a managed MCP control plane with a path to VPC or air-gapped deployment.

Its MCP surface covers more than remote proxying. TrueFoundry MCP Gateway supports a curated registry, standard OAuth flows, virtual MCP servers, OpenAPI-to-MCP conversion, hosted stdio MCP servers, and checks before and after tool calls. A virtual MCP server can present a selected set of tools from several underlying servers to one team or workflow.

That combination matters in a mixed tool estate. A company might have one remote HR server, one OpenAPI billing service, and one command-line data tool. TrueFoundry offers a route to govern all three without pretending they share the same hosting model. Enterprise supports full VPC and air-gapped installations for both control and gateway planes.

The wall is the step from an inexpensive prototype to enterprise-grade control. Developer is useful but limited. Pro is $499 per month. Pro Plus jumps to $2,999 per month, and Enterprise requires a quote. The higher published tier does not improve the price per included tool-call capacity: Pro is $0.499 per 1,000 included calls, while Pro Plus is about $0.600 per 1,000. The buyer is paying for controls, users, servers, and scale, not a volume discount.

Best for: Teams combining remote, OpenAPI, and hosted stdio MCP tools under private-ready governance
Standout: Virtual MCPs, OpenAPI conversion, hosted stdio, and pre-call plus post-call guardrails
Pricing: Developer $0, Pro $499 per month, Pro Plus $2,999 per month, Enterprise custom
Free trial: Seven days, no credit card required

Developer costs $0 per month for 50,000 requests, three users, up to five registered MCP servers, and 50,000 tool calls per month. Pro costs $499 per month for 1 million requests, 10 users, up to 25 MCP servers, and 1 million tool calls. A Pro overage block adds 2 million requests and five API keys for another $499 per month.

Pro Plus costs $2,999 per month for 1 million requests, 25 users, up to 50 MCP servers, and 5 million tool calls. Additional usage requires contacting sales. Enterprise uses custom pricing, targets 10 million-plus requests per month, and provides custom limits for users, MCP servers, and tool calls.

TrueFoundry beats Portkey when the tool formats and deployment location are the hard parts. Portkey wins when the servers are already running and the buyer wants a lower-cost authentication proxy. AgentCore wins when AWS-native function and API transformation plus broad managed identity are the center of the architecture.

5. Kong AI Gateway: best for teams already on Konnect

Kong AI Gateway is the best choice when Konnect already governs the company's APIs and the next task is extending that control to model, MCP, and agent-to-agent traffic.

Kong AI Gateway 3.14 made Agent Gateway generally available for production on April 13, 2026. The platform now presents three related boundaries: an LLM Gateway for model traffic, an MCP Gateway for tool and data access, and an Agent Gateway for A2A communication. Kong's release adds agent identity and authentication, unified observability across those traffic types, and inspection of agent-to-agent messages.

For a current Konnect customer, that is a credible consolidation. The same platform team can extend policies and operating practices from APIs to new agent traffic instead of procuring a separate gateway for every protocol. Plus permits unlimited MCP proxies and unlimited agents for A2A traffic, with up to five unique LLM models.

For a greenfield team, price is the wall. The cheapest Plus control plane is serverless at $25 per month, and each unique LLM model proxied through the paid AI plugins costs $100 per month. Five models plus one serverless control plane cost $525 per month before inference, extra API requests, bandwidth, or add-ons. Hybrid raises the control-plane line to $200 per month. Dedicated cloud is $500 per month plus $0.15 per GB.

Best for: Existing Kong customers extending API governance to LLM, MCP, and A2A traffic
Standout: One governance family across the complete AI data path
Pricing: 30-day trial $0; Plus bills per gateway and model; Enterprise custom and billed annually
Free trial: 30 days, no credit card required

The 30-day Free trial costs $0, requires no credit card, removes gateway limits, and keeps 30 days of analytics. Plus permits up to five serverless, two hybrid, and two dedicated cloud gateways. It includes 1 million API requests per month, charges $200 for each additional 1 million, and caps the tier at 10 million monthly requests.

Enterprise removes gateway and model-proxy limits, includes paid AI Gateway plugins and AI Gateway Manager, and uses custom annual pricing. That is appropriate for a broad platform commitment, but it prevents a public end-to-end budget.

Kong's decision rule is simple: existing platform value counts. If Konnect is already paid for, staffed, and approved, Agent Gateway can remove another vendor boundary. If not, the $525 five-model floor is paying for an API platform before the agent has proved it needs one.

6. Cloudflare AI Gateway and Agents: best low-fee edge stack

Cloudflare AI Gateway is the best low-platform-fee option for teams willing to combine its free model gateway with the separate Cloudflare Agents runtime and tool surface.

Cloudflare AI Gateway's core dashboard analytics, caching, rate limiting, and DLP scanning are free on all plans. Provider inference passes through without markup. Unified Billing is optional and adds a 5% fee to purchased credits, so $2,000 in credits creates a $100 billing line.

Managed agent tools live next door on Cloudflare Agents, a durable agent platform with Browser, Sandbox, AI Search, MCP, Payments, and other MCP tools. Its starters include server-side and client-side tools, human-in-the-loop approval, and task scheduling. The runtime can use OpenAI, Anthropic, Google Gemini, or other providers.

That is a capable stack, but it is not one unified product surface. AI Gateway governs model traffic. Agents hosts the loop and exposes tools. Workers and the selected tools carry their own usage and limits. A buyer who needs a single place to transform internal APIs, hold outbound OAuth credentials, and approve every tool call should not infer that the free gateway core supplies all of it.

Best for: Edge-native teams that want free gateway controls beside a managed agent runtime
Standout: Free core gateway plus Browser, Sandbox, AI Search, MCP, and Payments on Cloudflare Agents
Pricing: Gateway core free on all plans; Workers Free and Workers Paid have different log allowances; Unified Billing adds 5%
Free trial: Free core gateway and Workers Free allowance

Workers Free stores 100,000 persistent logs total across all gateways. Workers Paid stores 10 million logs per gateway. Logpush is available on Workers Paid and includes 10 million requests per month before $0.05 per additional million. Guardrail evaluation uses Workers AI token-based inference, so the free gateway does not mean every policy action is free.

Cloudflare wins when the team already builds on Workers and wants to keep model control, stateful agents, browser work, search, and MCP near the edge. It loses when one auditable tool-control plane matters more than low gateway cost. AgentCore and TrueFoundry make that boundary more explicit.

Who should pick what

Choose Amazon Bedrock AgentCore Gateway when the agent must call internal systems and the organization wants a managed service to transform, authenticate, discover, and invoke those tools. The choice flips to a smaller product when most tools are already exposed as stable MCP servers.

Choose Vercel AI Gateway when adding search is the work blocking a product release. The choice flips when the workflow must write to internal systems, support a mixed private tool estate, or maintain broad outbound OAuth connections.

Choose Portkey when the servers exist and every client should authenticate once. It is especially sensible for a platform team that needs a registry, workspace access, content filters, approvals, and logs without adopting a full agent runtime. The choice flips when the team still needs to build or host the servers.

Choose TrueFoundry when private deployment, hosted stdio, OpenAPI conversion, virtual MCPs, or pre-call and post-call guardrails carry the decision. The choice flips to Portkey for a lower-cost remote proxy or to AgentCore for an AWS-centered estate.

Choose Kong AI Gateway when Konnect is already the approved API control plane. Existing staff, policies, and vendor approval can make the incremental agent layer cheaper than the published greenfield price suggests. The choice flips when the team is buying Kong only for a small agent pilot.

Choose Cloudflare AI Gateway and Agents when Workers is already home and the team is comfortable composing the gateway, runtime, and tools. The choice flips when security or procurement needs one product surface and one owner for both model and tool traffic.

There is also a valid no new gateway answer. If one provider already supplies the model, tools, identity, logs, and budgets the workflow needs, another control layer creates a failure surface without removing work. Before buying, name the exact account, secret, adapter, policy, or pager the gateway makes disappear.

The tool-plane bill: why a free gateway can still cost more

Gateway pricing and tool pricing belong on different lines. The current search example makes that split unusually clear because AWS Web Search and Vercel Exa Search both list $7 per 1,000 queries.

For 100,000 monthly agent interactions, assume one semantic tool search, three tool invocations, 100 indexed tools, and one web-search query per interaction. AgentCore's gateway layer is $4.02: $2.50 for Search API calls, $1.50 for tool invocations, and $0.02 for indexing. Web Search adds $700. The combined visible lines are $704.02 before models, runtime, policy, observability, network, and backend services.

Vercel's token gateway has no platform fee. At 100,000 Exa searches after August 31, Exa adds $700. The visible combined lines are therefore $700 before models and any optional reporting or team-wide policy charges.

The $4.02 difference should not decide the architecture. AgentCore's example also prices semantic tool discovery and three internal tool calls for every interaction. Vercel's example prices the native search integration. The point is not that one is four dollars cheaper. The point is that both hide a $700 tool line behind a gateway with a tiny or zero platform fee.

The same check changes other rankings:

• Portkey: One million recorded requests cost $130 on Production before model and MCP-server hosting.
• TrueFoundry: Pro costs $499 for 1 million monthly tool calls; Pro Plus costs $2,999 for 5 million. The larger tier is not a unit discount.
• Kong: Five model proxies and the cheapest serverless control plane cost $525 before inference and add-ons.
• Cloudflare: The gateway core is free, but Unified Billing adds $100 to $2,000 of purchased credits, and Workers plus tool usage remain separate.

Use hard provider spend limits behind a gateway where the provider supports them. A gateway budget can stop traffic at the control layer, but it should not be the only ceiling on a credential that can also be used elsewhere.

The ones to avoid

Avoid Vercel when internal API transformation is the main job

Vercel's managed Exa integration is useful precisely because it is narrow. Do not stretch that evidence into a claim that the gateway will transform a private OpenAPI estate, hold every outbound credential, and govern every internal action. Choose AgentCore or TrueFoundry when those are the requirements.

Avoid Kong for a small greenfield pilot

Kong can be the correct enterprise platform and the wrong first experiment. Five model proxies plus the cheapest serverless Plus control plane cost $525 per month before inference. If the team has no Konnect estate to extend, begin with the workflow and prove its control needs before adopting the broader platform.

Avoid Cloudflare when one policy plane is mandatory

Cloudflare AI Gateway and Cloudflare Agents complement each other, but they remain separate surfaces. Do not let the free core gateway obscure the architecture review. Name which product owns model logs, tool execution, approvals, secrets, and incident response before calling it a unified control plane.

Avoid Portkey when nobody owns the MCP servers

A governed proxy cannot repair an unowned tool server. If a remote MCP endpoint fails, leaks scope, or changes its schema, Portkey can record the call but another team still has to restore the service. Choose a product with hosting or transformation when that is the missing work.

Avoid a paid tier until one action earns it

An agent with a wide tool catalog looks impressive and is difficult to authorize. Start with one reversible action, one owner, and one stop condition. The distinction between reusable agent instructions and callable tools also matters: agent skills and tool APIs solve different reuse problems. Do not buy a gateway to fix a definition problem in the architecture.

The Monday move: price one agent path end to end

Pick one workflow the business already pays people to complete, such as checking an order and updating a support ticket. Write the path as a sequence: user request, model call, tool discovery, credential exchange, tool invocation, result, approval, and audit record.

Give every step an owner and a monthly line. Include model tokens, gateway requests, search or browser calls, tool hosting, logs, security review, and incident response. Put temporary promotions in a separate column in the working sheet so they cannot become the permanent architecture assumption.

Then run one bounded path for a week. Keep the first action reversible. Set a provider ceiling, a gateway ceiling, and a stop condition for authorization failure, tool error, or unacceptable output. Review the logs with the owner of the system being changed, not only with the AI team.

The Monday decision is concrete:

• If internal adapter and identity work dominates, prototype AgentCore or TrueFoundry.
• If one search integration blocks shipping, use Vercel and budget $7 per 1,000 queries after August 31.
• If credential distribution dominates, place Portkey in front of the existing MCP servers.
• If platform consistency dominates, extend Kong or Cloudflare only where that estate already has an owner.
• If no named cost or control surface disappears, keep the current architecture.

The winning gateway is the one that removes a budget line or an operational owner while preserving the permissions of the system behind it. A feature that cannot be tied to that consequence is not a reason to migrate.

Frequently asked questions

What is an AI gateway?

An AI gateway is a control layer between an application and AI providers. It can add routing, retries, observability, policy, and spend controls. Some gateways now also govern tools and agent-to-agent traffic, but those capabilities should be verified separately from model routing.

What is an MCP gateway?

An MCP gateway controls how clients discover and call Model Context Protocol servers and tools. Its useful jobs include caller authentication, upstream credential handling, per-tool access, approval, logging, and sometimes tool transformation or hosting.

Which AI gateway is free?

Vercel includes $5 in monthly Free credit, Portkey has a Free Forever Developer tier, TrueFoundry has a $0 Developer tier and a seven-day trial, Kong offers a $0 30-day trial, and Cloudflare's core gateway is free on all plans. AgentCore Gateway is consumption priced, though new AWS customers may receive up to $200 in credits. Model, tool, runtime, log, and optional policy charges can still apply.

What is the best LLM gateway?

For managed agent tools, Amazon Bedrock AgentCore Gateway is the strongest overall. Vercel is better for zero-setup search, Portkey for access to existing MCP servers, TrueFoundry for private MCP control, Kong for a current Konnect estate, and Cloudflare for an edge-native stack. A buyer choosing only model routing should compare token fees, routing, caching, and provider controls separately.

Get the AI Business Workflow Audit Checklist

Turn one agent idea into a scoped workflow with an owner, budget, permission boundary, and stop condition. Subscribe to get the checklist free.