Automation & Voice

Beyond a single conversation, the agent has capabilities that persist and extend across sessions. For what it remembers between sessions, see Long-term memory.

Loops (recurring tasks)

The agent can re-run a prompt (or a wrapped skill) on a recurring trigger with the loop tools — create a loop, list active loops, cancel one, and persist small state in a per-loop scratchpad. Each iteration starts fresh. See the loop skill.

Example: "every 30 minutes, check this page and notify me if the price drops."

Triggers aren't only time-based. A loop's trigger can be time:<5m|1h|…>, time:dynamic, an event:<topic>, or page state: change (e.g. a selector on the current tab), combined with AND/OR. Its mode (chat / browser / agent) picks the tool catalog each iteration gets.

Two options make loops cheaper and self-checking (both new in v0.3.11):

• gate — a deterministic pre-check that runs before the agent each fire and skips the run entirely (no model cost) when nothing changed. An http gate GETs a URL and diffs an extracted value; a bash gate diffs a command's stdout; an event gate filters the trigger's payload.
• verify — a check over each iteration's tool results that records pass/fail on the run. It reuses the /goal check semantics (matches substring or /regex/, optional tool scoping, negate).

Self-improving loops. A loop can propose improvements to itself based on its recent runs (loop_evolve); you then accept or reject the proposal (loop_proposal_respond) before anything changes.

Loops vs. schedules. Loops are for fast, event-driven repetition (down to minutes). For calendar-style jobs at an hourly-or-slower cadence, use Scheduled tasks.

Web tools

The agent has web search and web fetch tools for pulling in information from the open web independently of the page you're on.

Voice (TTS / STT)

NevoFlux includes text-to-speech and transcription:

• ElevenLabs (cloud TTS), Kokoro (local TTS), and Whisper (speech-to-text). Configure voices/keys under the [tts.*] sections of config.toml.