Speedle for AI

Standalone public exploration website: https://ai.speedle.io

The site has two parts. Use AI to manage access covers the open-source Speedle policy modeler skill, the Speedle MCP launch preview and an AI-based risk-detection approach being built; risk detection is not presented as a shipped feature and never grants access that policy denies. Control what AI can access shows policy patterns across the AI lifecycle: model training, deployment, operations and inference, agents, MCP tool calls and agent-to-agent calls. These are solution patterns on the Speedle policy engine, not shipped platform integrations. The home page gives one sentence and one representative rule per area. The full SPDL examples, the longer explanations and the target AI risk-detection flow are on /patterns/ in English and /zh/patterns/ in Chinese, linked from the home page. Every SPDL example on the patterns pages is parsed and evaluated with the Speedle file store (eval.NewFromFile) before it is published; keep each SPDL block byte-identical to what was tested. The Chinese patterns page retains the same seven SPDL blocks, including their English comments. The zero standing privileges sample runs from samples/zsp/ and the delegated agent sample from samples/delegation/; neither is part of the sitemap.

The policy modeler is an open-source Agent Skill for any agent that supports the Agent Skills format. It lives in skills/speedle-policy-modeler/ and turns a handbook or regulation into SPDL, a design document and executable test cases. examples/en/ and examples/zh/ record two synthetic runs, each validated with spctl check and 37 authorization cases against a local PMS and ADS.

Ask your agent to install the skill from its GitHub directory. The skill README documents direct installation, prerequisites and usage for Claude Code and Codex. For a generic install, copy the skill folder into the agent's skills directory. The home page provides Claude Code and Codex install/use examples. OpenAI's official Build skills documentation documents Codex's $HOME/.agents/skills directory and $ skill invocation.

Install in Claude Code

Claude Code and Node.js 22 or later are required. Add this repository as a marketplace, then install the speedle plugin:

claude plugin marketplace add teramoby/ai.speedle.io
claude plugin install speedle@speedle

In Claude Code, configure the plugin:

/plugin configure speedle@speedle

Set the ADS endpoint (default http://127.0.0.1:6734) and service name. Set a bearer token only when the ADS requires one; the token is sensitive configuration and is stored in secure storage. The plugin provides the /speedle:speedle-policy-modeler skill and the plugin:speedle:check MCP server. The check server runs npx -y @speedle/[email protected] check and exposes only is_allowed by default.

Policy management is a separate opt-in plugin:

claude plugin install speedle-manage@speedle

Configure it with /plugin configure speedle-manage@speedle. It provides only the plugin:speedle-manage:manage MCP server and runs npx -y @speedle/[email protected] manage. Its PMS endpoint defaults to http://127.0.0.1:6733; select the speedle-cloud backend explicitly when using Speedle Cloud. Its token is also sensitive configuration. Never grant manage mode to an agent that is itself governed by the policies in that Speedle service. Use this plugin only for human-supervised policy authoring.

Third-party marketplaces do not auto-update by default. Refresh the marketplace and update the installed plugin:

claude plugin marketplace update speedle
claude plugin update speedle@speedle

Use claude plugin update speedle-manage@speedle to update the management plugin after refreshing the marketplace.

The interactive example uses simplified browser-local policy rules and does not simulate AI risk detection. It is not a Speedle engine, MCP integration, RAG adapter, or production service. The Speedle MCP section is a launch preview. @speedle/mcp is published on npm. The install and client configuration snippets match the speedle-mcp README: npm install @speedle/mcp, npx @speedle/mcp check and npx @speedle/mcp manage. @speedle/sdk is published on npm: npm install @speedle/sdk. Speedle AI uses white backgrounds with light-purple accents; the header, Products navigation, sphere favicons and footer follow the shared three-site format of speedle.io.

SEO and GEO

sitemap.xml, robots.txt, llms.txt and llms-full.txt at the root are hand-maintained; there is no generator. The sitemap and both home pages declare hreflang alternates for en, zh-CN and x-default. The complete Chinese site is at /zh/ and /zh/patterns/. Every factual claim in these files must already be true on the page: @speedle/mcp and @speedle/sdk are published on npm, Speedle Cloud is coming soon, and AI risk detection is not shipped.

node scripts/check-seo.mjs verifies that the sitemap and llms link targets resolve to files and anchors in this repository, that each sitemap entry's hreflang entries match its page, that every page's JSON-LD blocks parse and cover Organization, WebSite and SoftwareApplication, and that no page has a visible FAQ section or FAQPage data. It also checks the language-switch links, Chinese metadata and diagram labels, Cloudflare Web Analytics on every page, the shared header and footer, and byte-identical SPDL blocks on the patterns pages, including comments. For llms.txt and llms-full.txt it checks every external link for a direct 200 without a redirect, and keeps the canonical Speedle definition and the status claims (SDK version, npm, risk detection, example results) in sync with the page copy. No dependencies; the link check needs network access.

This is the beginning of the README. Read the rest at the original. Read the full README ↗