Talon ClawTalon Studio v1.0.2-beta

Chat

Chat with Talon Claw at /claw: the Talon Assistant by default, or any enabled agent, with streamed answers grounded in your evidence and intel.

Chatting with Claw

Open Talon Claw → Chat (/claw) for the full page with your thread list, or type into Ask Talon Claw in the top bar from anywhere. Ask a free-form question and the response streams token by token, interleaved with the tool calls the agent makes as it reads your evidence and intel, so you see the reasoning and not just the answer.

Agent picker. The composer's agent menu defaults to the Talon Assistant and lists every enabled agent, including custom ones, so you can ask a specialist directly.

Reasoning. With Extended reasoning on in Settings, thinking-capable models show a collapsible reasoning block.

Setup gaps. With no AI provider configured, the reply is an amber card linking to Talon Claw → Settings rather than an error.

Build automations with AI

When a request is really a repeatable workflow, open the Flow builder (/flow/automations) and use Build with AI: describe the automation and Talon Claw drafts the flow onto the canvas. It needs the engineer role in the workspace and a configured AI provider, and is rate-limited by FLOW_GENERATE_RATE_LIMIT (default 10 per minute per client IP). It only loads a validated definition onto the canvas for you to review; it never saves or activates the flow.

Context & history

Per user. Chat threads are yours, stored server-side per user. The active workspace is read at send time, so switching workspaces mid-conversation redirects the same thread's context; start a new thread when you want a clean slate.

Playbooks. Every built-in agent carries its own house procedure. Vendored skill playbooks are pinned into its context only if an admin selected them for that agent (see Skills).

Propose-only. Chat can read and draft, but every side effect is staged for review: no auto-merge, no auto-deploy. A read-only user gets the read tools only.

Limits. Chat is rate-limited by ASSISTANT_RATE_LIMIT (default 30 requests per minute per client IP) and refused with 429 once the org's token or cost budget is spent.

Voice

The composer has three voice controls: dictation, voice replies, and a conversation mode that combines them. Dictation runs on the server's CPU and replies are spoken by the kokoro container beside it, with no API key and no speech service involved.

Dictation

Speech-to-text with a local Whisper model (TALON_STT_MODEL, default onnx-community/whisper-base.en), up to 60 seconds per recording.

Voice replies

Reads replies aloud with the Kokoro-82M voice model, which runs in its own container, kokoro: the release compose file runs it and points TALON_TTS_URL at it, and Helm installs set config.ttsUrl to a Kokoro service they run. Each agent can have its own voice (set in the agent editor); the install default is TALON_TTS_VOICE, falling back to af_heart.

First use

The Whisper model is downloaded from Hugging Face the first time dictation is used and cached on disk; the Kokoro image carries its own weights. Until a model is ready, if the server has no route to Hugging Face, or with no Kokoro service, the browser's built-in speech recognition and voice are used instead; those are the browser's own features, and some browsers process speech on their vendor's servers.