Browse documentation
DocumentationRead Markdown

Current availability

Reviewed 2 November 2026 against application source 480760e. Source revision and rollout are separate checks.

Suzo is in private development. Public TestFlight, hosted signup, and shared hosted accounts are not open. These guides describe the latest development source; merging that source or updating this website does not update an installed iPhone app or a running service.

Match your versions

The development service and website were deployed on 2 November 2026, and the development iPhone was updated. Other installations still require their own matching service and native build; use the setup guides bundled with that service. Existing legacy contacts and receiver protocols remain supported; upgrading does not silently approve or replace a task session.

Development source behavior Guide
Existing tasks or explicitly approved new conversations per incoming call, with scoped consumers MCP setup
Native client exports, private plugin packages, status and discoverable guides Runtime integration
Durable progress/results, cancellation, and output after a speech window Call loop
Codex, native Hermes/Claude owners and custom runtime bridges Calling the agent
Shared helper builds, independent contact pins, migration and managed updates Local speech
Per-contact paged history, export/deletion, retention and diagnostics History and privacy
One-phone pairing, credential recovery, destination policy, and backups Self-hosting, security

Current native source stops local capture/playback when you hang up before waiting for the service. It retries pending push-token registration and separates foreground delivery from APNs rejection. The service can retain final provider usage after a call ends. Usage is operational accounting, not a provider invoice.

What has been checked

The source includes isolated HTTP/MCP, queue/ownership, package/export, helper lifecycle, native-state and history tests. A real ElevenLabs call on the development iPhone connected, carried spoken turns to the same Codex task, returned audible replies, ended, and retained its acknowledged transcript. That call exposed speaker routing and caption defects; the subsequent native fixes need another physical check.

The maintained Hermes owner completed native model admission, a human approval denial, and a scoped MCP call against an isolated test service. This does not certify every existing Hermes gateway or a physical callback. The guarded existing-session Hermes plugin has boundary tests and read-only native MCP connectivity checks; migration and physical callback validation remain outstanding. The Claude owner initialized the native SDK and discovered scoped MCP tools without a model turn; authenticated Claude calls remain unverified.

Portable plugin manifest/configuration schema checks and generated-file tests do not prove installation in a native client. Adapter protocol fixtures do not prove actual runtime admission or human approval routing. Local inference measurements do not establish real-room recognition quality. Physical locked/background ringing, audio routes, provider output, network recovery, login/reboot and sleep recovery remain separate validation.

Conversation and runtime support

Continue an existing task when its authorized owner is reachable, or deliberately approve a reusable profile that creates one fresh conversation for each incoming call. The latter approval survives receiver/PC restarts until revoked; its availability still expires without heartbeats. Existing-task and child credentials retain their expiry rules.

Common runtimes have configuration guides; custom runtimes use the same open MCP and callback contracts. Match each adapter to its supported runtime interface and selected policy. A second owner process is not automatic access to a live desktop chat. See harness capabilities and runtime integration.

Check without placing a call

Use get_suzo_status, get_suzo_guide, the app’s connection details and diagnostics, the service’s build/capability response, the intended session’s MCP tool discovery, and a receiver check where supported. A health response reports service/configuration state; configured provider keys or APNs settings do not establish delivery. Readiness checks must not place calls or wake a session.