Skybridge MCP Apps Framework

CI jobs generally use contents:read, generated view resources declare explicit CSP domains, installs use a frozen lockfile, and the manifest contains several transitive dependency security overrides; this supports moderate least-privilege and dependency-security scores. The README identifies server-model-UI data-flow topics, while tests expose tool input/result, model-context, download, and metadata channels, but the supplied files provide no complete policy for data classification, retention, telemetry, or credential handling. Applications may add connection/resource domains and view._meta may override framework-computed security metadata, so the boundary is configurable rather than strictly enforced. No user-confirmation mechanism before consequential tool calls is shown. External effects are application-defined, with examples involving booking, carts, OAuth, downloads, and model-context updates, but no uniform confirmation or side-effect annotation standard appears. Content hashes help isolate versions, yet no action undo or deployment rollback procedure is documented. The MIT file, author, named maintainers, organization, and security address provide attribution, but publisher identity remains unverified by the stated registry and customer-logo claims are not independently supported by the supplied material.

The README's type-safe MCP tools, React views, and multi-host claims align closely with the schemas, resource metadata, CSP construction, and CI/conformance workflows, earning strong self-consistency. Explicit Node and pnpm versions, frozen installs, Windows builds, and broad CI checks improve dependency availability; deductions remain because Notte, Slack, MCP hosts, and external hosting services are operational or verification dependencies without documented offline or degraded-mode guarantees. Tests demonstrate a specific duplicate-view exception, large-output warnings, failure artifacts, and notifications. The score is reduced because a missing production manifest silently falls back to an unversioned URI and the evidence does not show consistently actionable messages for all runtime failures.

The material explicitly serves both humans and coding agents and covers creation, migration, new views, commerce, travel, SaaS, authentication, and interactive UI scenarios, providing thorough audience coverage. Tool names and descriptions, Zod input/output schemas, optional views, host capabilities, and the one-tool-per-view rule provide useful capability boundaries and trigger signals. They are not complete enough for full scores because no repository-level agent manifest, unified trigger policy, or side-effect classification is shown. Environment fit is strongly supported by stated Claude, ChatGPT, VS Code, and compatible-client targets, Linux and Windows CI, example and documentation builds, a ChatGPT/Claude/Alpic conformance matrix, and self-hosting on Node-compatible platforms.

The README has clear sections for purpose, features, agent and human setup, documentation, deployment, community, maintainers, and categorized examples, supporting strong information architecture and examples. Installation commands are directly usable, but prerequisites and the full production procedure are deferred to external documentation not included here. Tests acknowledge a deprecated hosts option, and release branches plus a bump script suggest some migration discipline; however, the monorepo root is still version 0.0.0 and no changelog or stability policy is supplied. Known limitations are scattered through tests, such as one tool per view and manifest fallback, rather than collected for users. LICENSE and README say MIT while the root package.json says ISC, a material metadata contradiction that prevents a strong license score. Named maintainers, the team, issue and contribution channels, and a private security address clarify responsibility, but there are no response-time or support commitments and publisher identity is not registry-verified.

Structured content, text content, metadata, React views, CSP, development HMR, production asset hashing, and numerous examples make framework outputs directly usable for MCP apps, supporting full output-usability credit. Host abstraction, end-to-end type inference, and a local emulator plausibly add value over low-level SDKs and are partly corroborated by tests; deductions remain because claims such as seamless portability, the permanent tunnel, and the complete development experience are not fully evidenced in the supplied implementation excerpts. MIT licensing and self-hosting reduce adoption costs, while Node >=24.18, pnpm, browser testing, cloud conformance, authentication providers, and optional Alpic deployment add environment and operational costs. No performance or resource-cost measurements are supplied.

Many core claims trace to concrete tests and workflows: schema-based registration, resource URIs, content normalization, CSP, version hashing, host matrices, and cross-platform builds. README, test code, and workflow configuration provide useful cross-source corroboration, but this remains a static review without execution results, the lockfile, complete core implementation, or release history. Promotional assertions such as “seamlessly,” “delightful,” and “permanent tunnel,” along with customer-logo adoption claims, are not clearly separated from audited facts or supported inference, limiting fact/inference separation.