Accessor platform

Accessor is a local voice shell around harnesses (Codex, Claude Code, Antigravity, opencode, Cursor, mock). It owns the microphone, wake/sleep, speech, the dashboard, and routing. Each harness owns tools, sandbox, and its own conversation thread.

Audio overflow

Audio buffer overflow; restart Accessor before issuing more commands means the microphone callback filled a bounded queue faster than the DSP thread could drain it (CPU spike, a long burst, or a stall). That used to kill capture. Accessor now drops the overrun, resets the segmenter, and keeps listening. One occurrence is noisy, not fatal. Persistent repeats mean the machine cannot keep up — shorter turns, a quieter room, or fewer other CPU-heavy jobs.

Long speech is retained in overlapping 30-second chunks. Completed clips queue in order.

Wake vs conversation STT

Stage	Default	Optional
Wake word (29, hey 29)	Internal local STT (stt.engine, default Canary)	Always local — never cloud. Speech → Local STT model opens a list (Canary, Parakeet TDT 0.6B INT8, Whisper Tiny/Base/Small Q5). Esc leaves the list with no download. Only a missing model you then pick asks to confirm size.
After the session is awake (GREEN)	Internal local engine (stt.conversation=local)	External Cartesia Ink-2 (stt.conversation=cartesia)

/tts provider cartesia is spoken output (Sonic). /stt provider cartesia (aliases: ink-2, external) is after-wake transcription. /stt engine parakeet (or Speech → Local STT model) selects the on-device model. Wake spotting cannot use Ink-2.

Cartesia has two modes after GREEN. New profiles set stt.streaming=true, which takes effect only when Cartesia is selected; it uploads detected awake audio in ~100 ms chunks while the user speaks, before words/relevance are known. Turn streaming off to wait for the locally segmented utterance and a local word check before uploading the clip. Streaming requires stt.conversation=cartesia and a Cartesia API key. It is exposed under Speech → Cartesia streaming. Local endpointing sends finalize; only final transcript deltas acknowledged by flush_done become a prompt. Interim/partial results never trigger actions. Local transcription runs alongside the upload for wake controls and fallback. Wake-addressed commands do not wait for cloud results. Streaming is cancelled when capture is suppressed, the session sleeps, or speaker playback starts. Full local clips remain available after errors or overload; no automatic cloud retry replays a partial transcript.

Both modes retain 320 ms pre-roll, use a one-second silence endpoint, and chunk long utterances at 30 seconds with overlap. Noise can be mistaken for voice, especially in streaming mode; there is no claim that a VAD proves human intent. Only local wake detection runs during sleep/output. A rejected awake transcript appears in Activity with its reason, but is excluded from agent history and persisted analytics.

Compute saving (stt.lazy, default off): the microphone and Earshot VAD always run, and the segmenter already keeps 320ms of audio before the spike. The local STT session is the expensive part (Canary ~180MB; Parakeet INT8 ~640MB encoder). When this is on, weights load on the first ~160ms of voiced audio (usually while you are still talking) and drop ~45s after you go back to WHITE. First wake after unload can add a fraction of a second if the phrase is shorter than the load. ONNX Runtime itself stays in process; only the selected STT sessions are freed. Both Cartesia modes retain local transcription for controls and fallback.

Harnesses

The Main agent owns the conversation and defaults to Codex / gpt-5.6-luna / harness-default reasoning (currently low in coordinator mode). Coding and difficult analysis run in isolated workers, even when they use the same harness. Worker results return to main without copying their full tool logs. New profiles use Codex / gpt-5.6-luna for Accessor compaction. Explicitly configured providers are preserved; an unavailable compactor falls back to a bounded local excerpt and reports that fallback.

Settings → Harnesses lists Main agent first, then Coding agent and Compaction agent. Each opens three pages: harness → model → reasoning. Arrow keys and Enter work on every page; Esc goes back. All three values save together on the final page. Old routine configuration fields migrate to main without losing preferences.

Plugins & connectors is a preference passed to main, lower in the list. It defaults to Use main agent, following changes to main’s harness, model and reasoning automatically. Simple connector work can run directly in main; a different plugin preference or complex task can use a worker. Selecting the same harness never creates a worker merely because a plugin is involved.

Input relevance defaults to Jev when an existing TypeSafe key is available, with a local fallback otherwise. Jev receives the utterance and bounded recent conversation, and assesses whether it is addressed to the assistant and warrants a response/action. Explicit wake addressing establishes the first condition but does not force a response: a contextual dismissal such as “never mind” normally needs none. “Never mind that, stop the alarm” still requests an action. Rejected speech is not added to conversation history. The classifier runs asynchronously with a two-second timeout; outages fall back to local filtering. The local filter removes obvious filler/noise, not semantic intent. No classifier is called for silence or a bare wake.

A wake code interrupts output and active work, then opens listening. Intent is interpreted afterward. “Stop the alarm” reaches the main agent intact; the main agent uses stop_alarm and receives a receipt. Sleeping and typed Escape or /cancel remain immediate local controls. A bare wake provides at least eight seconds to begin speaking, stays silent, and holds automatic result delivery until a request or idle sleep. The usual conversation timeout remains 120 seconds.

The older routing.router / routing.auto-model settings still support legacy routing with the coordinator disabled; they are no longer part of the main settings flow.

Settings → Voice: think warble, wake chime, sleep chime, and the startup ready chime (0 silent, 1 normal; wake and sleep chime default to 1.5 so they carry over speech, and the think warble defaults to 1.5 on new profiles). A rising ready arpeggio plays once when Accessor has finished starting and is listening for the wake code. The warble follows unfinished work: it pauses during speech and resumes after progress speech, remaining active through tool calls until completion. Cancellation and pending approvals silence it. Status chrome — every box outline on the page — is green while the conversation is open, light blue while the agent works, and purple while TTS is speaking. Activity scrolls with the mouse wheel as well as PgUp/PgDn.

Caching, compaction, and usage limits

A warm process preserves a native conversation and avoids rebuilding it for every utterance. Provider prompt caching is separate: compatible requests can reuse computation for an unchanged input prefix. Keeping the same main harness/model and stable metaprompt helps; Accessor does not inject a changing timestamp into its system instructions. It cannot guarantee cache hits, expose a shared KV cache, or move cached state between vendors. A separate worker using the same provider/model may qualify for provider caching without sharing the main conversation, depending on the provider’s routing and cache rules. Model changes, different tool definitions, changed prefixes, compaction, and expiration can reduce reuse. The CLIs own cache controls and billing; Accessor does not claim savings without usage telemetry. See OpenAI prompt caching.

Compaction has two layers. Each CLI manages its native context; /compact (or the automatic threshold) separately summarizes Accessor-owned history through the selected Codex, Claude Code, Antigravity or Gateway provider, using its configured model and reasoning (Gateway effort support depends on the model). Explicit Local trim is available without inference. CLI compaction runs in an isolated process and cannot change main’s model. It is asynchronous, preserves turns arriving while it runs, and falls back to a bounded local excerpt when the provider fails. It reports this as local retention, rather than a model-generated summary. Cancellation stops its process. Brief harness handoffs use a bounded local excerpt so switching does not block voice controls. Notes, schedules and run receipts persist independently.

/limits reports observed provider failures and local backoff. Rate errors defer new work on that harness for five minutes; quota/usage exhaustion defers it for an hour. These are conservative local retry delays, not verified reset times or remaining subscription quotas. The errors and backoff persist in limits.json; scheduled work stays pending until eligible. Failed or uncertain actions are never automatically replayed. No paid fallback, credit purchase, or account switch occurs. Harness token/cost estimates in /analytics remain estimates.

Codex receives per-turn reasoning. Claude Code and Antigravity receive --effort and --model. opencode receives --variant and -m; Cursor receives --model. The main default effort is low; workers select low/medium/high explicitly. Claude/Antigravity keep one warm stdio process; opencode and Cursor expose only a one-shot headless invocation, so Accessor runs one process per turn and resumes the harness’s own session id (--session for opencode, --resume for Cursor) to keep continuity. Cancelling any of these terminates the current CLI process; the next main turn reconnects with an Accessor handoff. Worker cancellation and the 15-minute deadline stop the isolated worker and leave a receipt for scheduled work. Inspect incomplete external actions before retrying.

Fast mode (routing.fast-mode, off by default) is a best-effort low-latency path: it forces low effort and each harness’s light model for the main conversation and workers. It lowers latency and quality together and does not bypass any approval.

Notifications, watches, and quiet hours

A watch is a recurring scheduled task that carries survey guidelines. It runs as an isolated worker on Main’s harness and light model, may emit only notify controls, and has each finding scored by Jev against the guidelines; below-threshold findings are dropped and a Jev failure fails closed. Create, edit, list, and stop watches with /watch; the cadence minimum is 60 seconds. Findings become entries in notifications.json alongside agent-raised notifications (MCP notifications). The store is capped at 100, evicts read before unread, and suppresses an identical finding from the same source for six hours and a re-worded repeat for 72 hours. While the app is idle, unread items appear as Notice lines, ding with sounds.notify, and are spoken only when the raise asked for it and speak is on; each is marked announced so a restart does not repeat it. acc notifications list|read|read-all|dismiss and /notifications review them.

Quiet hours (night=HH:MM-HH:MM on a task or watch) defer a due task to the end of the window, wrapping past midnight; the deferral and a skipped-repeat count are recorded as run receipts and master-log events. Alarms are not deferred. Watches and scheduled tasks run only while an acc process is running. These are all local records; a watch finding is data for review, never authorization.

Approvals

Codex default is approvalsReviewer: auto_review with approvalPolicy: on-request. The Guardian/auto-review classifier handles sandbox escalations. Typed /approve N remains the human override. Claude uses its permission mode; Antigravity uses sandbox mode with plan/accept-edits according to workspace access. Accessor does not pass Antigravity’s skip-all-permissions flag. Worker approvals use typed /worker-approve N or /worker-deny N; voice cannot approve them.

Settings

Interactive dashboard: ↑/↓ move, Enter opens a category or toggles, Esc goes up/closes. The selected row shows a cyan description of what that setting does. Harness pickers list which CLIs are actually on PATH. Categories: Voice, Speech, Harnesses, Display, Metaprompt, Tests, Security, Computer. Typed /settings KEY VALUE still works for scripts.

Computer holds the desktop-control toggle plus the screenshot and zoom width caps. acc computer test captures a display without sending input; acc computer enable exposes the MCP computer tool to connected agents. It is off by default and is deliberately not agent-editable, so an agent cannot grant itself desktop control. The tool mirrors the standard computer-use action set and maps screenshot coordinates back to native pixels, and adds reliable semantic actions: open_app (launch by name), list_windows/focus_window, and ui_snapshot/ui_invoke/ui_set_value/ui_select/ui_expand, which drive apps through the platform accessibility tree instead of pixel clicks. The backend is Windows UI Automation, macOS Accessibility (System Events, after granting Accessibility permission), and Linux AT-SPI2 (via python3-pyatspi, with wmctrl for window focus). acc computer info reports which backend is available. The agent is steered to prefer these and use screenshots only as a fallback. See README for the full action list and platform caveats.

Display holds the chat view, the harness-announce toggle, the voice metaprompt editor, and the master log. The metaprompt (prompt, plus appended prompt_extra) is edited with /prompt — tone presets butler, warm, terse, plain, extra instructions that survive preset changes, or $EDITOR. It is not in the agent-editable setting set, so an agent cannot rewrite its own instructions. The master log (logging.enabled, off by default; logging.level info/debug; logging.max-mb size cap) writes one timestamped line per routing, agent, watch, notification, memory, organizer, MCP and speech event, rotating to <path>.1; /logs prints the path and tail. It never contains credentials or raw audio. Speech holds the spoken provider and after-wake STT, streaming, and the noise gate; Voice holds wake, barge-in, chimes and the ready chime; Security holds the password and auto-lock; Tests runs the microphone, transcription, voice, and connector checks.

What you need installed

1. Install acc: use a prebuilt binary from the Releases page (scripts/install.sh on Linux/Apple Silicon macOS, scripts/install.ps1 on Windows) or rebuild from a checkout with .\install-accessor.ps1 / cargo build --release --bin acc
2. acc or acc doctor — downloads ONNX Runtime + the selected local STT model if missing (the Piper engine and voice install automatically when Piper is the selected provider)
3. acc tts setup — Piper (default) / system / Kokoro / Cartesia
4. At least one harness login: Codex (acc agent), optional claude, optional agy
5. New profiles work with no cloud key: local Canary STT, local Piper TTS, and Codex compaction. Optional cloud keys: Cartesia (TTS and Ink-2 after-wake), TypeSafe (Jev relevance, with local fallback), and the Antigravity CLI for Antigravity compaction.
6. Optional desktop control: acc computer enable (then acc computer test), or enable it in the setup wizard.
7. acc -wakecode 29 speak

Wake-once conversation

Wake opens an idle follow-up window (120 seconds by default). With barge-ins enabled, clean continuations while thinking are captured normally; only audible speech/alarm interruption requires the wake code. Capture uses a one-second silence endpoint, overlapping 30-second long-speech chunks, and a FIFO of completed clips. Playback waits for capture, transcription and relevance processing, including after synthesis. Bare wake does not flush the following captured phrase. A rolling 2.4-second local recognition window checks every ~600 ms without waiting for utterance completion; inference time is additional. Say the wake code, pause, then give the request. Rolling probe text is never submitted as a prompt. /audio exposes check/hit/echo-rejection counts and decoding time. Sleep closes active listening and keeps the local wake detector; separate mute/unmute state is removed. AEC uses speech, thinking warble, alarm and chime playback reference audio, with residual text filtering and epoch checks. This reduces echo loops but does not identify the human speaker. Physical hardware testing is still needed.

Live MCP controls use a session-bound loopback capability and acknowledge only after the session applies sleep or stop_alarm. Detected Antigravity registration is checked on startup; Codex/Claude receive per-session MCP settings. Barge-in decoding skips long completed speaker clips, and stale wake mentions in old replies no longer veto interruption throughout subsequent speech. /audio exposes capture, VAD, windows, inference errors and the latest transient probe recognition.

Quotas and MCP preferences

/usage shows subscription bars and resets for all three harnesses; acc usage --json and MCP usage_status expose the same fields. Codex uses its account limits protocol, Claude its experimental get_usage control request, and Antigravity its standalone tab-separated /usage report. Queries are concurrent, bounded, and send no model turn. Missing data is unavailable; expired or older-than-five-minute samples are visibly stale. Status-line JSON can be ingested with acc usage --ingest claude|antigravity. /analytics remains approximate costs and latency; /limits remains observed-error retry delays. See README for provider caveats and schema sources.

MCP settings_read lists agent-editable keys; settings_update validates an atomic changes patch. Main’s live bridge applies supported preferences and returns an actual receipt. Harness changes take effect on the next turn; TTS changes on subsequent playback. Standalone/worker connections explicitly save for next launch. Settings tools do not expose credentials, approval policy, filesystem/executable paths, arbitrary metaprompts, the master log, or memory.capture. Plugin settings choose harness/model/reasoning and whether to follow main; they do not authorize accounts.

Shared memory (memory.json) is ranked by importance × 0.995^age_hours × (1 + ln(1 + access_count)); importance starts from the fact’s kind (explicit 1.0, identity 0.9, preference 0.7, fact 0.6, inferred 0.5) and access count rises at most hourly, so the injected digest favors facts that keep being useful without pinning stale ones. With memory.capture on (default), explicit, identity, and preference statements are captured as they arrive and a debounced pass every 15 minutes asks the configured compaction model for durable facts the heuristics miss, tagging them inferred and deduplicating near-restatements. /memory review lists live entries, /memory infer extracts from the last twelve turns, and /memory forget N drops one. Capture is a user-side privacy toggle, not an agent-editable setting.

MCP notes_search and note_read expose the user’s private Markdown notes as bounded context. note_delete permanently removes one exact search-result ID only after explicit user authorization; it has no bulk form. MCP initialization and the harness metaprompt advertise these capabilities. harness_health reports executable discovery and configured roles without treating discovery as proof of login. Organizer status includes device-local time and IANA timezone. Calendar schedules use local date/time and optionally every_days; they follow device-timezone changes and DST, while delay_seconds/every_seconds retain elapsed-time semantics.

tts.volume and sounds.alarm now have independent controls, alongside sounds.think, sounds.wake, sounds.sleep, sounds.notify and tts.speed (0.6–2.5; Cartesia is clamped to 1.5). Accepted transcripts have only a You entry; rejected awake input includes the ignored words and reason. Cartesia STT and Cartesia TTS are the live streaming paths; Piper, Kokoro, and system voices synthesize completed WAV chunks before playback.