Simon Willison — 2026-07-31#

Highlight#

The rollout of the new stateless Model Context Protocol (MCP 2.0) has drawn Simon back to the specification as a safer, more auditable alternative to giving AI agents arbitrary shell access. His enthusiasm translated immediately into code, resulting in three new open-source tools this week: a CLI explorer, a Datasette endpoint, and a client plugin for his LLM utility.

Posts#

Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp) · simonwillison.net Simon delves into the Model Context Protocol 2.0, noting its shift to a stateless, single-HTTP-request model that dramatically simplifies implementation compared to the legacy session-based approach. Having previously warned about the risks of giving agents arbitrary shell and internet access, he praises MCP tools for being easier to audit, control, and drive with local models. To explore the new spec, he rapidly built three tools: mcp-explorer (a Python CLI for inspecting servers), datasette-mcp (adding SQL capabilities to Datasette), and an alpha llm-mcp-client.

smevals - a small eval suite for evaluating models, prompts, and harnesses · simonwillison.net After years of iterating on evaluation frameworks, Simon collaborated with Prime Radiant to release smevals, a suite for running tasks across different configurations and grading the results. The tool formalizes a clear vocabulary for evals—distinguishing tasks, configs, runs, and checkers—and deliberately separates the execution of a run from the grading process. It also includes a command to compile the results into static HTML reports, making it easy to share benchmark outcomes.

llm 0.32rc1 · simonwillison.net This release candidate finalizes a major schema change for the llm CLI, introducing content-addressable hash IDs for stored messages. This enables database de-duplication and allows the tool to accurately represent complex, forked conversation trees generated by modern model families. Users are strongly advised to back up their logs.db before upgrading to this new structure.

llm-chat-completions-server 0.1a0 · simonwillison.net Leveraging the new de-duplicating schema in LLM 0.32rc1, Simon built a plugin that exposes all of a user’s installed local and remote models via an OpenAI-compatible Chat Completions endpoint. The server handles conversations where each incoming message extends the previous state, efficiently storing the growing payloads. Impressively, GPT-5.6 Sol wrote the entire plugin, demonstrating its strong grasp of the OpenAI API shape.

deepseek-ai/DeepSeek-V4-Flash-0731 · simonwillison.net Simon highlights DeepSeek’s new 304 billion parameter V4 model, noting its exceptional value-per-intelligence and its ability to outrank larger models like MiniMax M3. During his standard SVG generation test (a pelican riding a bicycle), the default reasoning level disappointed, but bumping the reasoning effort to “high” yielded a vastly superior output.

Oxide and Friends: The Open Weight Revolution with Simon Willison · simonwillison.net Simon joined Bryan Cantrill and Adam Leventhal’s podcast to discuss a wild week in the AI ecosystem, spanning the impressive capabilities of open-weight models like Kimi K3 and high-profile cybersecurity incidents. Alongside various historical digressions, they revisited their January predictions, with Simon adding a new one: the Pope will make a statement about open models by the end of the year.

Project Pulse#

Simon’s recent output reveals a concerted push to build robust, auditable infrastructure for working with LLMs. The synchronized schema changes in his llm CLI tooling, the release of the smevals testing framework, and the rapid adoption of Stateless MCP all point toward an ecosystem designed to safely integrate and evaluate AI capabilities.