FAQ

Answers to the questions that come up most. For step-by-step setup see Getting started; for problems see Troubleshooting.

Is Accessor free, and does it need API keys?

The default path needs no key and no account. Speech recognition (Canary) and the default voice (Piper) run locally, and the harnesses use their existing CLI logins. Cartesia (cloud speech in both directions) is opt-in and needs a key; Codex uses your existing login and Accessor never calls the OpenAI API directly.

Does my audio leave my computer?

Not by default. Wake detection, conversation recognition, and the default voice are local. If you select Cartesia for after-wake STT, detected awake speech is uploaded while you speak (only the final transcript is submitted to the agent), and wake detection and spoken unlocking stay local. While the password lock is active, cloud STT is disabled. See Features → Speech.

Which agents are supported?

Codex App Server, Claude Code, Antigravity (agy), opencode, and Cursor (cursor-agent), plus a mock harness that needs no account. See Harnesses.

Do I need Python?

No, not for the default path. Canary, Piper, and the gateway run natively in Rust. Python is only needed for the optional Kokoro voice worker and a few helper scripts.

What are the system requirements?

Windows x64 is the tested platform; Linux x64/ARM64, Apple Silicon macOS, and Windows ARM64 use cross-platform libraries but need real-hardware testing. Linux builds need the desktop dev packages; see Getting started → Platform prerequisites. Intel macOS is source-only.

How is this different from other local voice tools?

Accessor is a gateway, not another transcription app: it keeps the voice pipeline local, manages wake/sleep, routing, memory, and playback, and drives one or more existing CLI coding agents without replacing their tools, sandbox, or login. It reuses the agent’s own connectors rather than shipping a second OAuth store.

Can it run as a background service or on a headless server?

No. This release does not install a service; alarms, tasks, and watches are checked only while an acc process is running, and a powered-off or suspended machine is not woken. It is a desktop audio application, not a server build.

Does it work without a microphone?

Yes. acc --text --agent mock runs the full conversation plumbing with no microphone and no account, which is the fastest way to test or demo. --text never opens a microphone and is silent unless --speak is explicitly passed.

How do I change the wake word?

acc config set wake-code 29, or say/type a settings change in the dashboard. “Hey 29”, “hi 29”, and “ok 29” are accepted variants. The wake word reduces accidental activations; it is not authentication — see the local access lock.

Where is my data stored?

Settings and local state live in the directory shown by acc config path: config.json/device.json, memory.json, notes/, schedules.json, notifications.json, the optional master log, and the encrypted journal. Secrets go in the OS credential store, never in config.json. The full table is in Troubleshooting → Logs and data locations.

Can I move my settings to another machine?

Yes. acc config export FILE writes the portable subset and acc config import FILE merges it, keeping the target’s device settings. Copying the whole folder also works but carries device.json; delete it on the new machine or import instead.

What is the dictation journal?

An encrypted, local-only log of speech you explicitly dictate. It never goes to an agent or cloud service and is encrypted at rest with XChaCha20-Poly1305, with its key in the OS credential store. See Local-only dictation.

Can an agent control my computer?

Only if you enable it. Desktop control is off by default and is deliberately outside the agent-editable settings surface, so an agent cannot turn it on for itself. When enabled, it can drive the real mouse, keyboard, and screen — it is not a sandbox. See Desktop control.

How fast is it?

It depends on the model, microphone, and CPU. Accessor reports median/P95 timings in /analytics and ships repeatable benchmarks; see Voice performance. Local TTS uses completed WAV chunks, while Cartesia streams PCM.

How do I update Accessor?

Re-run the installer; acc update only updates the agent CLIs, not Accessor itself. See Releases.

Why do macOS and Windows warn about the app?

The binaries are unsigned (and macOS builds are unnotarized), so Gatekeeper and SmartScreen warn once. The workarounds are in Getting started. Code signing is planned but not shipped.

How do I report a bug or suggest a feature?

Use the issue templates. For anything involving the access lock or a privacy boundary, read Local access lock first and describe the impact privately rather than in a public issue.