{"id":50976858,"url":"https://github.com/teaglebuilt/skill-cards","last_synced_at":"2026-06-19T09:02:17.296Z","repository":{"id":364817505,"uuid":"1269278986","full_name":"teaglebuilt/skill-cards","owner":"teaglebuilt","description":null,"archived":false,"fork":false,"pushed_at":"2026-06-14T15:21:31.000Z","size":1583,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"master","last_synced_at":"2026-06-14T17:16:19.918Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":null,"language":"TypeScript","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/teaglebuilt.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":"CONTRIBUTING.md","funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-06-14T14:16:33.000Z","updated_at":"2026-06-14T15:21:35.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/teaglebuilt/skill-cards","commit_stats":null,"previous_names":["teaglebuilt/skill-cards"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/teaglebuilt/skill-cards","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/teaglebuilt%2Fskill-cards","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/teaglebuilt%2Fskill-cards/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/teaglebuilt%2Fskill-cards/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/teaglebuilt%2Fskill-cards/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/teaglebuilt","download_url":"https://codeload.github.com/teaglebuilt/skill-cards/tar.gz/refs/heads/master","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/teaglebuilt%2Fskill-cards/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34523991,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-19T02:00:06.005Z","response_time":61,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-06-19T09:02:16.392Z","updated_at":"2026-06-19T09:02:17.287Z","avatar_url":"https://github.com/teaglebuilt.png","language":"TypeScript","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Skill Cards\n\nA demo of portable, sandboxed tools for AI agents — built by composing two orthogonal extensions of the Model Context Protocol.\n\nThis repo is the working argument behind a thesis about the new design space MCP's extension framework enables. Two example skills, a sandboxed Wasm runtime, a stateless MCP server, and a TanStack Start web app that proves the host side works outside the existing major MCP clients.\n\n![Skill Card Example](./docs/assets/consent-component.png)\n\n## What\n\nA **skill card** is a single OCI artifact bundling four things: a WebAssembly Component (the logic), a WIT interface (the typed contract), one or more UI resources (HTML surfaces the user interacts with), and a permission manifest (capabilities the component needs, with rationales). Author once, push to a registry, and any host that implements MCP Apps and a Wasm runtime can pull it down, execute it sandboxed, and render its UI inline in the conversation. Closer to a browser tab than installing a cli.\n\n## How\n\nThe architecture is three layers stacked on the MCP base protocol, each owning a distinct concern.\n\n**Distribution** uses OCI registries. Each skill card is a content-addressable artifact in `application/wasm` + `application/json` + `text/html` parts, pushed by `oras` and pulled by Wassette or the loader's local file reader. The OCI manifest hash is the version identity — a skill upgrade is just a new tag pointing at a new manifest. Authoritative content, no separate package metadata.\n\n**Execution** uses WebAssembly Components running in Wasmtime. Each call to a skill's tool builds a fresh `WASIShim` configured from the consented permissions for that specific invocation. The component sees a sandbox containing only what the user agreed to expose: a single preopened directory, a specific outbound host, no environment variables it wasn't granted. Tool functions are declared in WIT and exposed as MCP tools with auto-derived JSON schemas. The execution surface is provided by `@bytecodealliance/jco` in this demo and would swap to Microsoft's Wassette in production without code changes — same WASIShim contract.\n\n**Presentation** uses MCP's UI resource mechanism. When a tool returns a result, it can declare `_meta.ui.resourceUri` pointing at a UI resource declared in the same manifest. The host fetches the resource, renders it in a sandboxed iframe attached to the result, and brokers `postMessage` communication between the iframe and the host. UI-initiated actions flow back through the same MCP protocol as direct tool calls — same audit, same consent path.\n\n## Why\n\n**Sandboxing is per-call, not per-tool.** A skill card's manifest declares `fs:read` with a `{argument:path}` template; the `WASIShim` built for an invocation grants access only to the specific path the agent passed. Granting \"always allow\" once doesn't expose the entire home directory on future calls — it caches the *resolved* permissions for that specific scope. Calling the SBOM auditor on `~/proj/foo/package-lock.json` doesn't authorize calling it on `~/proj/bar/package-lock.json`. This is what makes the policy comprehensible to humans: you're never granting an abstract capability, you're granting specific access to a specific thing for a specific stated reason.\n\n## Further Documentation\n\n1. [CONTRIBUTING.md](./CONTRIBUTING.md)\n2. [Architecture](./docs/architecture.md)","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fteaglebuilt%2Fskill-cards","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fteaglebuilt%2Fskill-cards","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fteaglebuilt%2Fskill-cards/lists"}