OpenCode Mobile

The README clearly describes the HTTP/SSE flow among the mobile client, the user's self-hosted opencode server, and the model provider, and says provider keys do not pass through the app's servers. Credentials are described as OS-keystore protected, crash reporting as opt-in and off by default, and tool calls as reviewable and rejectable; biometric protection can also cover individual sends. Deductions apply because the supplied files do not show the Android permission manifest, production approval enforcement, certificate-validation policy, or complete retention behavior. package.json includes PostHog and the build injects a PostHog key, creating tension with the statements that no analytics SDK is bundled and no personal data is collected. Dependencies are generally constrained and CI builds a source-oriented F-Droid variant, but no vulnerability scan, update policy, or lockfile evidence is supplied. Diff review and rejection can prevent changes, yet no post-change rollback or recovery mechanism is shown. Independence, upstream opencode, Expo, corporate copyright, and MIT attribution are explicit; unknown registry provenance was not treated as suspicious.

CI defines type checking, unit tests, Android builds, and mock-backed activation E2E coverage, including a visible 401 failure and several directory, diff, and picker flows. Server prerequisites, Android scope, and installation channels are also documented. Self-consistency is reduced because the README repeatedly identifies 0.4.7 as current while package.json reports 0.4.15, and the no-analytics claim conflicts with the PostHog dependency and build key. Workflow comments acknowledge that newer flows are non-blocking and had not completed in CI, while the streamed-reply assertion is omitted because of a harness limitation. The authentication-negative flow supports a useful failure-message score, but the evidence does not cover user-facing handling for broader timeout, disconnection, or server-version failures.

The intended audience and scenarios are unusually clear: Android users connect to a self-hosted server through LAN, Tailscale, Cloudflare Tunnel, or ngrok, with an offline demo for evaluation. Stable, beta, planned, and unavailable features are separated, and the client is accurately bounded as a thin interface rather than a model provider or hosted backend. Approval prompts, connection testing, and biometric send protection indicate reasonable trigger precision, but the production enforcement code is absent, so uniform gating of high-impact actions cannot be confirmed. Environment support includes multiple connections, directory scoping, and several distribution channels, but it remains Android-only, tablet layout and offline history are unfinished, and no real-server compatibility matrix is provided.

The README has strong information architecture covering installation, quick start, architecture, status, roadmap, privacy, contribution, support, and licensing. Three Android installation routes and server startup instructions are adequate for ordinary onboarding. iOS absence, beta tunnel setup, and planned tablet and offline-history features are stated, while the MIT text and copyright are complete. Deductions apply for unstable metadata: the README says 0.4.7 but the package says 0.4.15, and the @opencode-ai/mobile package name may blur the otherwise explicit non-affiliation statement. Examples and an offline demo exist, but there is no systematic FAQ or troubleshooting section. CI can generate release notes, yet no changelog or migration guidance is supplied. A maintaining company, sponsorship route, support email, and project tracking are named, but curated publisher identity is unverified and no clear security-response ownership or support commitment is shown.

Streaming chat, resumable sessions, inline diffs, approval controls, and cross-directory session management produce directly usable mobile coding interactions. The offline demo also lowers evaluation effort, giving the product meaningful value beyond a generic remote shell. Deductions reflect its dependence on a user-operated and secured opencode server, network exposure or tunneling, and separately billed model access; the tunnel wizard is beta, while offline history and managed hosting remain unshipped. The client is free and has several install options, but total cost and benefit still depend on server, provider, and secure-network setup.

Several claims are traceable across the supplied files: package.json identifies dependencies and scripts, workflows specify tests, builds, signing, and the F-Droid flavor, and the mock server combines detailed protocol commentary with concrete endpoint behavior. Traceability is reduced by the version and analytics/privacy contradictions and by the absence of privacy implementation code, permission manifests, a lockfile, release attestations, or actual test results. Features, builds, and test intentions receive useful cross-file corroboration, but not every promotional claim is independently supported. The material carefully distinguishes the mock from a real server, blocking from non-blocking tests, harness limitations, beta functionality, and future plans, so fact, inference, and aspiration are separated well.