If you use Claude only through the chat window at claude.ai, you have seen one face of it. The other face is the Claude Platform, which Anthropic has also called the Claude Developer Platform. It bundles the Claude API, the client SDKs, and the Claude Console, the web interface that moved from console.anthropic.com to platform.claude.com when Anthropic unified its developer offerings under the Claude brand. It is where developers, ISVs, and technical teams build products and integrations on top of Claude, rather than typing prompts one at a time.
This matters more than usual right now. Salesforce and Anthropic recently announced an expanded partnership, bringing Claude deeper into Salesforce's own product surface. That announcement covers the customer-facing side: a plugin, prebuilt skills, tighter integration across Claude, Salesforce, and Slack. This article covers the other half, the Claude Platform infrastructure underneath any Claude-Salesforce integration, whether it comes from Salesforce itself, from Anthropic directly, or from a third-party MCP server your team already connects.
For Salesforce admins, consulting partners, and anyone evaluating AI tooling for their org, understanding this layer matters. It is the foundation that AI clients, integrations, and Salesforce MCP servers connect through. Knowing what it offers helps you evaluate any AI vendor that claims to work with Claude, including the ones you connect to Salesforce.
What the Claude Platform actually is
The Claude Platform is Anthropic's developer surface for building with Claude. At its core sits the Claude API, a REST API at api.anthropic.com that gives programmatic access to Claude models. Developers send requests, get structured responses, and build applications around them, including the Salesforce integrations covered below.
The Platform gives you a few core pieces:
- Direct model access. You construct each turn, manage conversation state, and build your own tool-calling loop. This is the traditional way to integrate Claude into an application.
- Managed agent infrastructure. Instead of wiring up your own loop, you can deploy agents into stateful sessions that keep event history and run autonomously against defined tools, environments, and success criteria. Anthropic calls this Claude Managed Agents, and it is still in beta.
- The Console itself. A web interface where you generate API keys, test prompts in a browser playground, organize work into workspaces, and manage billing and usage.
Feature availability can vary depending on how you access Claude. Anthropic-operated surfaces include the Claude API, Claude Platform on AWS, and Microsoft Foundry. Partner-operated surfaces include Amazon Bedrock, run by AWS, and Google Cloud, run by Google. Each has its own rollout timeline for new capabilities.
The core building blocks
Underneath the Console, a handful of concepts show up in almost every integration built on the Platform. Understanding them helps explain what any AI vendor means when they say their product "runs on Claude."
Messages and conversation state. Every request to the API is a list of messages, alternating between user and assistant turns, plus an optional system prompt that sets context and instructions. There is no separate "system" role inside the message list itself. Because the API has no memory between calls, the calling application is responsible for sending the full conversation history each time it wants Claude to have context from earlier turns.
Thinking. For complex problems, Claude can reason through the task before producing a final answer, and that reasoning comes back in thinking blocks ahead of the response. On current models this is adaptive: Claude decides when and how much to think, and developers steer the depth with an effort setting rather than the fixed token budget older models used. The thinking text you can see is a summary, never the raw chain of thought, and the newest models hide it by default until you opt in. Thinking tokens are billed as output and count against the same token limit as the response itself. This is what lets Claude handle multi-step problems, like reconciling conflicting data or planning a sequence of tool calls, more reliably.
Tool use. Claude can call external tools mid-conversation, whether that is a web search, a code execution sandbox, a browser, or a custom function a developer defines. The Platform supports parallel tool calls, meaning Claude can request several tools at once rather than waiting on each one sequentially. This is the same mechanism behind a Claude API Salesforce MCP server: the server exposes Salesforce actions as callable tools, and Claude decides which ones a request needs.
Structured outputs and citations. Structured outputs let a developer specify a JSON schema and get a guaranteed schema-conformant response back, useful for anything that needs to plug directly into another system without extra parsing. Citations let Claude interleave source references directly into its output when working from provided documents or search results. The two features cannot be used together in the same request, since citations require freeform text interleaving that a strict JSON schema does not allow.
Prompt caching and batch processing. Repeated context, like a long system prompt or a set of reference documents, can be cached for 5 minutes or up to 1 hour, cutting cost and latency on subsequent calls that reuse it. For work that does not need an immediate response, the Message Batches API processes large volumes of requests asynchronously at a lower cost than real-time calls, which matters for anything like bulk document analysis or large-scale data classification.
Workspaces and key types. Inside the Console, workspaces separate environments, teams, or use cases, each with its own spend controls and usage tracking. API keys come in different types, including personal keys and service account keys, so organizations can track usage and manage access at a more granular level than a single shared key. An Admin API lets larger organizations manage these resources programmatically instead of through the Console UI alone.
What changed recently on the Claude Developer Platform
Anthropic has been shipping quickly on the Platform side. A few developments stand out for anyone building or evaluating AI tooling right now.
Computer Use, Skills API, and Files API reached general availability. These three capabilities moved from beta to production-ready status together. Computer Use lets Claude operate a graphical interface directly, reading screenshots and interacting with on-screen elements, now with support for multiple sequential actions in a single model call. A new browser tool adds webpage parsing for web-based agent work. The Skills API lets teams package domain expertise and organizational workflows into reusable folders Claude can load when relevant. The Files API got a meaningful capacity increase, with higher rate limits and expanded storage per organization.
Managed agents gained more governance controls. Teams running agents through the Platform can now set session budgets, choose advisor models, pin inference to a specific geography for data residency requirements, and pull in GitHub-hosted skills. These are the kinds of controls that matter once an AI tool moves from a proof of concept to something running against production systems.
Self-hosted environments entered public beta, on the Claude Code side. This one sits next to the Platform rather than inside it: organizations on Claude Team and Enterprise plans can now run Claude Code cloud sessions on infrastructure they control. The organization builds and maintains a runner image, and either keeps a fixed set of runners running or hosts an orchestrator that starts runners on demand as sessions queue. It is off by default, is not available to organizations with zero data retention enabled, and is aimed at teams whose network or compliance requirements a fully managed environment cannot satisfy.
Security scanning extended to skills and plugins. Also adjacent to the Platform: Claude Enterprise organizations can turn on automatic scanning of third-party skills and plugins for malicious content at upload or edit time, across Claude, Claude Cowork, and Enterprise plugin marketplaces. As more teams pull in community-built or vendor-built skills, this closes a real gap.
Pricing held steady. Claude Sonnet 5's introductory pricing became the standard price, with a previously scheduled increase called off. For teams budgeting API spend, that is one less variable to plan around.
Mid-conversation tool changes moved into beta. Developers can now add or remove tools between turns of an ongoing conversation without losing the benefit of prompt caching, using a dedicated beta header. Previously, changing the available tool set mid-conversation could force a full cache miss, adding cost and latency. This matters for any agent that needs to adapt which tools it can call as a task evolves, rather than fixing its tool set at the start.
Response headers got more precise for multi-workspace organizations. The API now returns a workspace ID header on every response, identifying exactly which workspace an API key or access token resolved to. For organizations running several workspaces to separate teams, environments, or client engagements, this makes usage attribution and auditing more exact.
Structured outputs and citations are stable across the current model lineup. Structured outputs, which guarantee schema-conformant JSON responses, are supported on every current model and see real production use for anything that pipes Claude's output into another system without manual parsing. Citations, which let Claude interleave sourced references directly into a response, work alongside prompt caching: the citation blocks themselves are not cached, but the source documents they reference can be, so citing from a large document set does not mean paying full price for that context on every call.
Why this matters if you are evaluating AI tools for Salesforce
Most Salesforce teams will never touch the Claude API directly. But the Claude Platform is the foundation underneath the AI clients and integrations you might already use, including Claude, Claude Code, and the growing ecosystem of MCP servers that connect AI tools to business systems, Salesforce included.
When a vendor tells you their tool "works with Claude," the Platform is usually part of the answer to how. The governance features Anthropic ships here, like permission scoping, workspace separation, and security scanning, set the baseline that responsible integrations build on top of. Cirra AI is one of the products built to work across Claude and other MCP-compatible AI clients, connecting them to a Salesforce org so admins and consulting partners can manage metadata using natural language instead of clicking through Setup menus.
A few practical questions follow from this, worth asking any AI vendor pitching a Salesforce integration:
- Does it run through the standard API, or a managed agent? A managed agent with persistent session history behaves differently under load and over long-running tasks than a simple request-response integration. Neither is automatically better, but the answer affects how the tool behaves during a long admin session or a multi-step configuration change.
- How does it handle permissions? Nothing in the Platform itself enforces Salesforce-specific permission logic. That has to be built by whoever connects Claude to the org, typically through OAuth scoped to the authenticated user. Ask how the vendor scopes access, not just whether they claim to.
- What happens to your data? Prompt caching, structured outputs, and other features each have different data retention behavior, and some are eligible for zero data retention while others are not. If your org has compliance requirements, this is worth asking about directly rather than assuming.
- Which model, and why? The Platform supports a family of models with different capability and cost tradeoffs. A vendor's choice of model, and whether they let you choose, says something about how they have tuned the integration for cost versus reliability.
That is a reasonable starting checklist for evaluating any AI tool for Salesforce that runs on Claude, whether the integration comes from Salesforce's own product surface, a specialist Salesforce MCP server, or a custom build from an implementation partner.
Whether or not you ever open the Claude Platform yourself, it is worth knowing it exists and what it does. It explains a lot about how the AI tools already in your stack got built, and what is likely coming next.
For a deeper look at Salesforce's own Claude partnership announcement, see Claudeforce, Explained for Salesforce Admins.
