Harnesses

A harness is a command-line coding agent that Accessor drives. Accessor owns the microphone, wake/sleep, speech, the dashboard, and routing; the harness owns its tools, sandbox, permissions, and its own native conversation thread.

Supported harnesses

Id	CLI	Notes
codex	Codex App Server	Default on Windows; uses your existing login. Accessor never calls the OpenAI API directly
claude	Claude Code	Uses the installed CLI and its login
antigravity	agy	Passes the active workspace to its server
opencode	opencode	Read headless; management commands are interactive
cursor	cursor-agent	Read headless; reads .cursor/mcp.json
mock	none	No account needed; useful for testing the plumbing

A “gateway” and “local” compaction path also exist for summarization without a harness. Model discovery reads the selected Codex installation’s account catalog without running an LLM; the last successful list is cached for offline display.

Roles

Three roles can use the same or different harnesses and models:

• Main — the persistent conversation, with a lightweight model by default.
• Coding — isolated workers for code and difficult analysis.
• Plugin — connector/plugin work; defaults to following Main.

acc config set routing.main claude
acc config set routing.main-model haiku
acc config set routing.coding codex
acc config set model gpt-5.6-sol

The metaprompt tells the main agent to delegate coding, difficult analysis, and plugin work — even when all roles use the same CLI. A worker gets a fresh process, its own model and reasoning level, a 15-minute deadline, and returns a bounded result to the main conversation. Only one worker runs at a time, and worker controls cannot recursively delegate. See Routing and delegation.

Models and reasoning

Each harness→model→reasoning combination is selectable in /settings → Harnesses. Say “switch to Codex” or “switch model to Astra” using a name from your discovered list. A running harness can also emit ACCESSOR_SWITCH harness=codex (or JSON accessor_switch) to pin this conversation to another CLI without rewriting the plugin/coding/main slots. Accessor passes the switch through the same wake-code and barge-in rules.

Fast mode (routing.fast-mode) is a best-effort low-latency path: it forces low effort and each harness’s light model for the main conversation and workers.

Connectors are reused

Accessor does not implement a second Gmail client, OAuth store, or connector SDK. Configure connectors in the agent itself:

acc connectors setup
acc connectors list
acc connectors status
acc connectors refresh

• Codex opens its native UI for setup and asks App Server for accessible apps.
• Antigravity uses its native plugin catalog and MCP registry.
• opencode / Cursor have interactive management, so Accessor reads their own configuration (opencode.json mcp/plugin, .cursor/mcp.json mcpServers) and injects the cached inventory into each launch.

Restart Accessor after changing integrations. Status reports configuration/runtime availability, not a completed external tool call.

Accessor’s MCP server

acc mcp serves Accessor’s own tools over stdio. Codex and Claude launched by Accessor receive a session-specific connection; existing connectors remain available. acc mcp-install repairs registration manually.

Tool	Purpose
memory	Search and maintain shared facts
notes_search / note_read / note_delete	Private Markdown notes (delete needs an exact ID and user authorization)
organizer_control	Notes, alarms, tasks, sleep, stop_alarm, list_schedules
notifications	List, read, dismiss, read-all
harness_health	Executable discovery, configured roles, workspace, cached model count, timezone
usage_status	Subscription quota observations (cached by default)
settings_read / settings_update	Narrow, validated preference surface
session_control	Sleep and ringing-alarm stop through an authenticated loopback bridge
delegate_task	Start an isolated worker; the result returns asynchronously

Registration preserves unrelated entries and refuses name collisions. Native harness tool permissions still apply. Session capabilities are supplied to main sessions, not to isolated workers or compaction jobs.

Permissions and approvals

The default Codex sandbox is read-only; --workspace-write --workspace PATH permits workspace edits under Codex’s policy. Structured command/file approvals and empty-form MCP confirmations require a typed /approve N and expire after 60 seconds. Authentication and device-verification requests, forms requiring field values, and permission-profile grants are declined rather than guessed — complete those in the harness’s native interface.

See Features → Permissions and scope.