Claude model APIs and tool use.
Publisher documentation ReviewedTHE INFRASTRUCTURE BEHIND THE AGENT
The AI agent stack.
From intent to action.
A model reasons. A stack gets things done. Explore the layers that let agents remember, use tools, communicate and transact—with the controls to run them.
THE MAP
What are you building?
An architectural map, not a shopping list. Products can span layers; a small agent may need only a few of them.
01
Think & understand
Models, state and the context to make decisions.
Reason
Models & routing
Anthropic
OpenRouterCoordinate
Frameworks & orchestration
LangGraph
MastraRemember
Memory & retrieval
Mem0
QdrantFind
Search & external knowledge
Exa
Firecrawl02
Act in the world
Tools, execution, communication and transactions.
Connect
APIs, tools & MCP
ApideckMCP ConnectExecute
Browsers & sandboxes
Browserbase
E2BCommunicate
Email & communications
AgentMail
TwilioTransact
Payments & commerce
AgentCard by Alchemy
Stripe03
Run with control
Identity, reliability, evaluation and user experience.
Authorize
Identity & authorization
Auth0
WorkOSRecover
Durable execution
Temporal
InngestImprove
Observability & evaluation
Langfuse
BraintrustDeliver
User interfaces & voice
AI SDK
LiveKitAcross every layerScoped accessHuman approvalsData boundariesAuditability
HOW THE PIECES CONNECT
Tools still do real work.
An MCP connection can wrap an existing API. It does not replace that API or grant new permissions.
- Agent runtime
- MCP client & server
- Provider API / resource
- Result back to the agent
EXPLORE THE LAYERS
A purpose for every component.
01Models & routingGenerate responses, select tools and route inference requests.
Design decisionCompare tool-calling behavior, latency, data handling and model fallback policies.
A common API for model and provider routing.
Publisher documentation Reviewed02Frameworks & orchestrationTurn model calls into stateful loops, tool decisions and handoffs.
Design decisionChoose explicit state, stopping conditions and human handoff behavior before adding multiple agents.
Graph-based orchestration for stateful agents.
Publisher documentation ReviewedTypeScript agents, workflows and application services.
Publisher documentation Reviewed03Memory & retrievalPersist useful context and retrieve relevant records across runs.
Design decisionSeparate conversation history, long-term memory and source documents; define deletion and tenant boundaries.
Long-term memory across agent sessions.
Publisher documentation ReviewedVector search and metadata filtering for retrieval.
Publisher documentation Reviewed04Search & external knowledgeBring current web evidence and page content into the agent context.
Design decisionRetain source URLs and freshness; retrieved content is input, not permission to execute instructions.
Web search and content retrieval for agents.
Publisher documentation ReviewedWebsite scraping, crawling and content extraction.
Publisher documentation Reviewed05APIs, tools & MCPGive agents typed operations over business systems and external services.
Design decisionMCP standardizes a tool connection; it does not replace the underlying API, its scopes or its access rules.
Unified APIs for business-software integrations.
Publisher documentation ReviewedA library for discovering, connecting and dispatching MCP tools.
Publisher documentation Reviewed06Browsers & sandboxesRun code or interact with websites in a controlled execution environment.
Design decisionScope filesystem, network and credential access; choose APIs when a direct operation is available.
Cloud browsers for web automation.
Publisher documentation ReviewedIsolated environments for running agent-generated code.
Publisher documentation Reviewed07Email & communicationsGive an agent an inbox or messaging channel to receive requests and deliver results.
Design decisionAn email address is a communication endpoint, not proof of identity or permission to act for its owner.
Create inboxes and send, receive and thread email through an API.
Publisher documentation ReviewedProgrammable messaging channels such as SMS and WhatsApp.
Publisher documentation Reviewed08Payments & commerceConnect approved purchases, checkout and payment operations to agent workflows.
Design decisionKeep payer authorization, spending limits, identity checks and merchant support explicit.
AgentCard by Alchemy
Request a payment for a specific purchase through its CLI or API.
Publisher documentation ReviewedPayment infrastructure and agent integration tools.
Publisher documentation Reviewed09Identity & authorizationEstablish who is acting, on whose behalf, and which resources they can access.
Design decisionUse scoped credentials, tenant isolation and explicit approval for consequential actions.
User authentication and delegated agent authorization.
Publisher documentation ReviewedApplication authentication and enterprise identity services.
Publisher documentation Reviewed10Durable executionPersist workflow progress through retries, waiting periods and process failures.
Design decisionPlan idempotency and recovery for external side effects; retrying must not duplicate a payment or message.
Durable workflow execution with recovery after failures.
Publisher documentation ReviewedEvent-driven durable functions and workflows.
Publisher documentation Reviewed11Observability & evaluationTrace model/tool behavior and evaluate outcomes against representative tasks.
Design decisionMeasure task success, latency, cost and failure modes; redact secrets and personal data from traces.
LLM tracing, prompt management and evaluation.
Publisher documentation ReviewedLogs, datasets and evaluations for AI applications.
Publisher documentation Reviewed12User interfaces & voiceExpose the agent through a usable chat, application or realtime voice experience.
Design decisionMake progress, cancellation, approval and human escalation visible to the person using the agent.
Streaming AI application interfaces and model integrations.
Publisher documentation ReviewedRealtime voice agents and their application interfaces.
Publisher documentation ReviewedCONNECTION STANDARDS
Protocols connect the layers.
These have different jobs. They are not interchangeable, and listing a product here does not claim support for every protocol.
Understand the stack.
What is an AI agent stack?
An AI agent stack combines a model with orchestration, context and tools, plus the infrastructure needed to run it. Depending on the task, that can include memory, browser or code execution, identity, messaging, payments, durable workflows, evaluation and a user interface.
Does every agent need all twelve layers?
No. These are responsibilities to consider, not twelve mandatory vendors. A read-only research assistant may need no payments or browser automation. A framework or hosted service can cover several responsibilities.
Where do AgentMail and AgentCard fit?
AgentMail belongs in email and communications. AgentCard by Alchemy belongs in payments and commerce. An agent may need both communication and payment capabilities, but this map does not claim that every product pair integrates automatically. AgentCard by Alchemy at agentcard.ai is kept separate from the similarly named agentcard.com catalog profile.
Does MCP replace APIs?
No. MCP provides a standard interface between an application and tools or context. An MCP server can call a provider API, access a database or work with local resources. Provider authentication, permissions and limits still apply.
Is this a verified compatibility or popularity ranking?
No. The examples are an editorial selection classified by documented purpose. Publisher documentation was reviewed on the displayed date; inclusion does not establish interoperability, runtime testing, security certification or popularity.
How we map the ecosystem
This is an editorial architecture map from the Apideck team, not an exhaustive vendor inventory or a ranking. Examples are grouped by documented purpose; products may span several layers. Sources and review dates are shown with each example. We have not runtime-tested these combinations.
Apideck and MCP Connect are built by the team behind API Tracker. AgentCard by Alchemy is linked to its publisher documentation: the existing similarly named agentcard.com profile represents a different identity.
Hosting, storage, secrets, deployment and incident response also need an implementation plan. The use-case highlights are illustrative architecture choices, not preconfigured integrations.