APIs as Skills via MCP
Your agents can now consume web APIs as Skills via MCP. This is the foundation of the BoltMCP preview release, and solves many of the context problems people have experienced with MCP implementations. It also addresses the security challenges associated with giving agents access to APIs via CLIs or other code execution.
With this feature you can serve multiple APIs from different vendors under a single MCP server.
Even when the underlying API specifications amount to 10,000,000s of tokens, connecting to the server only costs ~1,000 tokens upfront.
Progressive disclosure 🤝 dynamic tool registration
Exposing entire APIs to agents is a context disclosure challenge, which we solve via MCP Skills. By implementing this pattern of in-context retrieval we remove the reliance on brittle RAG pipelines or band-aid solutions like client-side tool search.
Traditional Skills are great for providing context and template code to agents, but they don't provide ready-to-use tools. With APIs as Skills via MCP, we're adding tools to the equation.
Instead of preloading all tools into context, agents are empowered to discover what tools are available from connected APIs. Then, according to user privileges and consent, the agent can dynamically load the tools needed for the task. All tokens pulled into context are useful tokens.
No code mode, no sandboxes.
Using MCP as a mechanism for executing agent-generated code is a powerful pattern for managing context in complex workflows. However for interaction with web APIs we've found sandbox environments to be increasingly redundant, whilst also introducing unnecessary complexity with regards to security and infrastructure.
With a few token-efficiency tricks, traditional tool calling yields high performance whilst keeping access control and observability straightforward.
Native inline auth
Self-hosting your own MCP servers means authentication is implemented natively through your company's identity provider. Your MCP client handles OAuth tokens just like any other OAuth app, isolated from its agents.
Access control is enforced inline with OPA policies, autogenerated for each server and fully customizable. Any API keys or upstream OAuth tokens are handled entirely by the remote MCP server, never even touching the machine your agent is running on.
Tool approval without relying on the agent harness
Different agents implement their own tool approval patterns, often allowing users to create allowlists of tools which don't require approval.
With BoltMCP, the server itself enforces your approval and access control policies inline, which the agent cannot bypass. This dramatically reduces the risks involved when connecting your agents to your APIs.