Release notes
← Home · Installation · npm package
October 28, 2026 · published on npm
npm 2.10.3: experimental native conversion
ios-agent-mcp 2.10.3 brings the Universal App Engine foundation to the npm package. Inspect supported local source, produce a traceable App IR, and generate an editable native SwiftUI project.
What’s included
- Bounded HTML, React/TSX, static Vue and Svelte importers. Unsupported syntax is rejected explicitly; imported application code is never executed.
- React literal props and defaults, local primitive constants, typed syntax, and supported state, events and flat memory routes.
- Shared horizontal rows with conservative native sizing checks.
- Evidence comparison and resumable verification, with separate reference approval and baseline-promotion authorization.
Try the source-to-native workflow
npx -y ios-agent-mcp@latest app-ir inspect ./index.html MyApp > app-ir.json
npx -y ios-agent-mcp@latest app-ir generate ./app-ir.json ./MyAppRequires Node.js 20 or later. Native build and simulator verification need macOS and Xcode. Generation creates a new project; it does not by itself prove that the app works.
Verification and limits
Release checks: 828 tests passed, zero failures, one skipped. Package installation and conversion smoke checks passed. Specific fixtures compiled with Xcode 27; the row fixture was launched and visually inspected on an iPhone 18 Pro simulator. These results do not establish complete visual fidelity, accessibility, or arbitrary framework support.
The Universal App Engine remains experimental. General Next.js, Flutter and other framework conversion and autonomous repair are not complete. Studio remains a repository application, not an npm-hosted service.
Six compatible dependency updates were applied. Two high-severity npm audit findings remain for Zod and the simulator dependency path; npm reports no fix within the current constraints.
Supported inputs and architecture → · Inspect the native row evidence →
October 15, 2026 · published on npm
npm 2.10.2: design workflows and project readiness
ios-agent-mcp 2.10.2 was published on October 15. This release packages plan-driven design, synthetic sample data, screenshot-aware refinement, generated tests, iPad captures, remote macOS builds, provider modules and approval-gated shipping workflows.
It also includes deployment-target and scope-aware availability checks, local SDK/compiler evidence, and target-aware App Store readiness reports. Dependencies were updated to address npm audit findings.
Release validation: 582 tests passed, zero failures, one opt-in native test skipped; npm audit reported zero vulnerabilities. Registry version, latest tag and tarball integrity were verified. Live provider integrations and signed App Store distribution still require separate verification.
October 13, 2026 · available on GitHub main · subsequently included in npm 2.10.2
Deployment checks, readiness review and community discovery
The existing availability analyzer now reads deployment targets, checks guards in their enclosing scope and recognizes declaration annotations. Local SDK and compiler evidence is labelled separately from static fallback results.
App Store readiness now reviews target-specific permission strings, privacy-manifest evidence, entitlements, launch configuration and metadata. Ship preparation includes a hashed readiness report; it does not guarantee App Store approval.
Validation at implementation: 582 tests passed, zero failures; native SDK checks passed separately. Complex project configurations report coverage limits explicitly.
The Community & Mentions page now groups verified links with search and filters, including the Gemini CLI extension gallery. Read the readiness change · Read the deployment guard change.
October 3, 2026 · published on npm
npm 2.10.1: repair-path and CLI safety fixes
ios-agent-mcp 2.10.1 was published on October 3. This patch normalizes Xcode diagnostic paths across macOS workspace aliases so the repair loop can locate project files, bundles CLI 0.4.1 with cleanup/help safety fixes, updates vulnerable dependencies, and corrects installation instructions.
Validation: all 493 MCP tests passed, including the new path regression; npm audit reported zero vulnerabilities at release.
October 2, 2026 · on main · npm 2.10.0 and CLI 0.4.1 subsequently published
Audit: what was replaced, and why
After the agent landed, the whole repository and this website were checked for older code and pages that were wrong, out of date or duplicated. Each was replaced, and each fix has a test where code changed.
- The build budget no longer runs out before you build. The agent’s 25-minute limit used to start when the plan was written, so reading PLAN.md first could use it up. It now starts at the first build. Each refinement now gets its own budget; before, a project could never be built again after 8 attempts.
ios-agent clean --helpno longer deletes files. It used to run a real clean, and so did a mistyped option. Now every command prints help for--help, unknown options are refused, andcleanwill not erase a build-agent project’s history.--helpno longer runs commands. Before,install-command --helpinstalled a file andpreflight --helpran checks.loop statusalso stopped creating empty folders.- Correct install steps.
/ios-buildpointed at the npm version that does not contain the agent, and said XcodeGen was required. It now gives the working checkout install, and XcodeGen is optional. An old CLI pin (0.2.0) is now 0.4.0. - A steadier demo test. The Reading List UI test sometimes typed before the screen was ready. It now waits for each step, always uses the same simulator, and saves screenshots when it fails.
- One place to choose a workflow. Three guides overlapped on “turn an idea into an app”. One page now picks between the agent, the starter and the loop, and the others point to it.
- Out-of-date statements removed. The package page said version 2.8.0, that XcodeGen was required and that apps could not be generated. A setup page said integration tools were not on npm, though they shipped in 2.8.0. Xcode 27 pages said the test Mac had Xcode 26.6. Each now states what is true and what is still untested.
The new agent page, with the example apps → · Full changelog →
October 2, 2026 · on main · repository 3.9.0 and npm 2.10.0 prepared, not yet released
Build an app from one sentence
The new /ios-build command for Claude Code plans an iOS app (screens, data model, every capability it needs and the third-party costs), writes the SwiftUI code and Xcode project, then builds, runs and screenshots it in the simulator, ending with a report of what needs your accounts or money. XcodeGen is optional: the agent can write the Xcode project itself.
Capabilities are modules: 41 of them (Sign in with Apple, SwiftData, StoreKit, MapKit, Swift Charts, widgets and Live Activities, Apple Pay, passkeys and more), out of a catalog of 161. All 41 compiled together with Xcode 27.0 in Swift 6 mode. Four apps the agent planned and wrote, including a food-delivery app with 20 modules, built on the first build and ran in the simulator.
Not yet verified: the agent's own automatic build-and-fix loop on a Mac, each module's individual verify run, and purchases or Live Activities at runtime. The existing loop command keeps its separate job: making an existing project pass your own acceptance tests.
Example apps and screenshots → · How the agent works and what is verified →
September 30, 2026 · repository v3.8.1 · npm 2.9.0
Local Apple Developer Documentation Grounding
Use the Apple documentation installed with Xcode through its official MCP documentation tool. Inspect installed versions, retrieve bounded API evidence and consult SDK declarations when documentation access is unavailable. No Apple archive is bundled. Dynamic Xcode tool discovery and a doctor command now expose setup and capability evidence without opening or modifying a project.
Shared agent instructions and the bounded build-repair loop consult documentation for supported public API errors. Retrieval helps review availability; it does not replace compiling and testing. Bridge access requires Xcode approval, offline behavior is controlled by Xcode, and symbol-level SDK diffs remain unsupported.
September 29, 2026 · repository v3.7.0 · npm 2.8.0
Apple environment and optional AI providers
Inspect installed SDKs and simulators, select an optional text-generation provider, and keep local-only privacy as the default. Cloud access and fallback require explicit settings. Private handoff files and saved advisory role proposals preserve context; they do not autonomously edit apps or claim test success.
Screenshot Studio adds verified Mac, Apple TV, Vision Pro and Watch pixel profiles. Xcode 27-only APIs and Duo poses remain unverified on our Xcode 26.6 machine.
September 29, 2026 · repository v3.6.4 · npm publication deferred
App Store Screenshot Studio
Turn existing captures into localized screenshot sets with eight reusable layouts, neutral device frames, contact sheets and validated PNG exports. Nine new tools bring the source server to 56 tools; npm remains 2.7.2 with 40.
Copy and translations are supplied. Rendering is offline. Dimension and geometry checks do not imply Apple approval.
September 29, 2026 · repository v3.6.3 · GitHub source only
Apple system integrations
Ten integration workflows and seven new MCP tools provide Swift previews, evidence-based recommendations and separate permission/capability checks. Source builds expose 47 tools. npm remains on 2.7.2 with 40 tools until the next package release.
Scaffolding does not overwrite files. Device authorization, hardware and account behavior require testing in your app.
September 22, 2026 · npm 2.7.2 · repository v3.6.2
App colors and backend guidance, together.
One connection now exposes 40 tools. This update adds semantic palette generation, color review and backend integration review while preserving the existing app, simulator and release-preparation tools.
Build a color system
Generate 35 semantic roles across light, dark and high-contrast appearances, inspect contrast evidence and review catalog usage. Generation returns a preview; it does not silently edit your app.
Color generation and verification →
Review your backend boundary
Detect Supabase, Firebase, CloudKit, AWS Amplify, Appwrite, REST, Apollo/GraphQL and WebSockets. Review concrete credential, policy and networking evidence without sending your project to a provider.
What is verified
308 MCP tests pass. The backend Swift sample passes 9 offline tests and iOS Simulator-target typechecking. Login examples still need each app’s provider configuration; live authentication and deployed policies have not been verified. Historical benchmark results are unchanged.
September 21, 2026 · npm 2.7.1 · repository v3.6.1
Inspect the app. Prepare the evidence.
This patch adds conservative launch-screen reviews and the first local release-preparation workflow. It helps you identify missing project facts before attempting a release.
Launch-screen review
The existing App Store audit and project analysis now inspect launch configuration, storyboard and asset references using selected-target evidence. Thirteen diagnostics have regression coverage; unresolved settings remain explicit coverage gaps.
Checks and supported configurations →
Local release preparation
Use apple analyze to inspect a target and apple prepare to create an integrity-hashed local draft. The output separates observed facts, inferred candidates and unanswered questions. Unchanged inputs reuse a verified package; changed evidence creates a new package.
npx -y ios-agent-mcp@latest apple analyze --project /path/to/app --target App --configuration Release --json
npx -y ios-agent-mcp@latest apple prepare --project /path/to/app --target App --configuration Release --jsonThis phase does not connect to Apple, run an AI provider, archive, sign, upload or submit an app. Draft evidence is not a release approval.
Workflow, evidence states and limitations →
Fixes and verification
- Corrected JSONDecoder/JSONEncoder formatter false positives.
- 248 MCP tests pass, including 29 launch tests and 52 release-preparation tests.
- Twelve launch benchmark fixtures include valid and broken configurations, diagnosis/fix checks and unnecessary-change detection. Agent trials for these tasks have not been run.
The previous benchmark stays unchanged: 120 runs, 60/60 passes in each arm, and 60 paired ties. We do not claim measured improvement from this update.
Private reporting remains experimental
The 37th unified tool, private_feedback, requires a configured HTTPS receiver and explicit preview approval. Hosted delivery is not live by default. No silent public submission is added.
Next milestone
Evaluate the harder launch-screen tasks, inspect diagnosis quality and unnecessary changes, and resolve unsupported project configurations before extending release automation.