Stateless MCP has recaptured my interest

21 hours ago

Stateless MCP, also known as MCP 2.0 or the 2026-07-28 Model Context Protocol specification, represents the most significant update since its launch, reigniting interest in the protocol.
MCP provides a standard way to expose tools to LLM agents, addressing the risks of giving agents terminal and curl access by offering easier audit and control, suitable for smaller models.
The new stateless specification simplifies both client and server implementations, using a single HTTP request instead of the previous two-step initialization and call process, improving scalability by eliminating server-side session state.
mcp-explorer is a new stateless Python CLI tool for interacting with MCP servers, offering commands like list, inspect, and call, built with the help of Codex.
datasette-mcp is a Datasette plugin adding a /-/mcp endpoint, providing three tools: list_databases, get_database_schema, and execute_sql (read-only), enabling agents to query Datasette instances.
llm-mcp-client is an alpha plugin for LLM tool, integrating MCP and demonstrating the ability to run SQL queries against a Datasette instance, with plans for future integration into LLM core.
The author now views MCP as a safer approach for building with agents compared to general agents with shell and curl access, due to easier reasoning about capabilities and potential risks.

MCP 2026-07-28 introduces a stateless core, shifting from bidirectional stateful to request/response for serverless and edge deployment.
Standardized extensions for MCP Apps and Tasks provide a formal path to add capabilities like interactive UIs and long-running work without core protocol changes.
Authorization is hardened with production OAuth 2.0 and OIDC support, enabling integration with enterprise identity systems like Entra or Okta.

MCP (Model Context Protocol) is a promising development for connecting tools in an LLM-agnostic way, but real-world support varies significantly among major providers.
Anthropic (Claude) offers seamless MCP integration out of the box, but its API is expensive and token consumption can be high for simple use cases.
OpenAI provides remote MCP support but the author encountered a 'Session terminated' error and could not get it working, indicating it needs more testing.

WorkOS launches a remote MCP server that lets AI agents manage WorkOS accounts via OAuth, inheriting dashboard permissions.
The server exposes hundreds of operations across WorkOS products (organizations, SSO, Directory Sync, AuthKit, etc.) but excludes credential-minting, impersonation, billing, and secret fields.
It uses a discover-then-execute design with four tools (whoami, list_operations, query, mutate) to save context window space.

The 2026-07-28 Model Context Protocol (MCP) specification introduces a stateless protocol core, transforming MCP from a bidirectional stateful protocol into a request/response stateless one for improved reliability and scalability.
Every request is self-describing with optional discovery; methods and tool names travel in HTTP headers for routing and authorization.
Multi Round-Trip Requests (MRTR) replace server-initiated bidirectional streams, enabling stateless interactions like user input requests mid-call.

MCP (Model Context Protocol) has significant token consumption issues, using up to 15,000 tokens for a single tool like Playwright, which limits usable context window.
LLMs face non-linear computational scaling with context length, making long sessions expensive and slow.
Selectively disabling MCP tools is not supported in some platforms like Claude Code, leading to frequent context length limits.

MCP 2026-07-28 introduces a stateless core, shifting from bidirectional stateful to request/response for serverless and edge deployment.
Standardized extensions for MCP Apps and Tasks provide a formal path to add capabilities like interactive UIs and long-running work without core protocol changes.
Authorization is hardened with production OAuth 2.0 and OIDC support, enabling integration with enterprise identity systems like Entra or Okta.

MCP (Model Context Protocol) is a promising development for connecting tools in an LLM-agnostic way, but real-world support varies significantly among major providers.
Anthropic (Claude) offers seamless MCP integration out of the box, but its API is expensive and token consumption can be high for simple use cases.
OpenAI provides remote MCP support but the author encountered a 'Session terminated' error and could not get it working, indicating it needs more testing.

WorkOS launches a remote MCP server that lets AI agents manage WorkOS accounts via OAuth, inheriting dashboard permissions.
The server exposes hundreds of operations across WorkOS products (organizations, SSO, Directory Sync, AuthKit, etc.) but excludes credential-minting, impersonation, billing, and secret fields.
It uses a discover-then-execute design with four tools (whoami, list_operations, query, mutate) to save context window space.

The 2026-07-28 Model Context Protocol (MCP) specification introduces a stateless protocol core, transforming MCP from a bidirectional stateful protocol into a request/response stateless one for improved reliability and scalability.
Every request is self-describing with optional discovery; methods and tool names travel in HTTP headers for routing and authorization.
Multi Round-Trip Requests (MRTR) replace server-initiated bidirectional streams, enabling stateless interactions like user input requests mid-call.

MCP (Model Context Protocol) has significant token consumption issues, using up to 15,000 tokens for a single tool like Playwright, which limits usable context window.
LLMs face non-linear computational scaling with context length, making long sessions expensive and slow.
Selectively disabling MCP tools is not supported in some platforms like Claude Code, leading to frequent context length limits.

Hasty Briefsbeta