Configure the AI providers every agent run depends on, set org-wide behaviour and budgets, and watch token and cost usage, all at /connect/ai.
Providers
The Providers & behaviour tab of Talon Claw → Settings (/connect/ai) holds an ordered list of AI connections. The first enabled global connection is Primary, later ones are fallbacks, and a workspace-scoped connection overrides the global chain for its workspace. Add provider offers Anthropic, OpenAI and xAI; presets for OpenAI-compatible services (Ollama, OpenRouter, Groq, DeepSeek, Together AI, Fireworks AI, Mistral, Google Gemini); and Self-hosted (custom) for any other OpenAI-compatible endpoint. Claude Code (subscription) runs calls through the Claude Code engine, billed to a Claude subscription rather than a metered key: the token from claude setup-token, entered on the connection or set as CLAUDE_CODE_OAUTH_TOKEN on the server.
Scope
Global (all workspaces), or a single workspace.
Purpose
Chat (conversation and assistant): the agent brain. Embeddings (semantic search): optional, since a local model works out of the box. Tool-calling for agents (optional): a model with reliable function calling for agent runs, used when the Chat model narrates tool calls instead of making them; agents fall back to the Chat model when unset.
Model
Fetch available models pulls the live list from the provider. Before you fetch, Talon suggests Claude Opus 5, Claude Sonnet 5, Claude Opus 4.8 and Claude Haiku 4.5; GPT-4o and GPT-4o mini; or Grok 3 and Grok 3 Mini. For a self-hosted endpoint, enter the base URL first, then fetch the server's list.
API key
Required for the native providers and presets; optional for a self-hosted endpoint. Encrypted at rest and never shown again.
Base URL
Self-hosted only: the endpoint including its /v1 path (a bare host gets /v1 added).
Allow self-signed TLS certificate
Self-hosted only: skips certificate checks for that endpoint. Prefer trusting your CA with NODE_EXTRA_CA_CERTS.
Verified on add. Adding a provider runs a real test call and discards the connection if the test fails, so a bad key or URL never lands in the fallback chain. Saved cards carry Connected, Failed or Untested badges, with Test connection in each card's menu.
Reorderable fallback.Move up and Move down set priority; disable or delete a card and the next enabled connection takes over.
Admin-only, with guards for self-hosted models
Listing, adding, testing and changing connections, and every Behaviour setting, need the admin role. The appliance also guards self-hosted endpoints: a private or localhost base URL (including Ollama on the same host) needs TALON_ALLOW_PRIVATE_AI_BASE_URLS=true, and self-signed TLS needs TALON_ALLOW_INSECURE_AI_TLS=true.
Behaviour & budgets
The Behaviour panel sets org-wide options; each saves as soon as you change it:
Extended reasoning
Lets thinking-capable models show a collapsible reasoning block. Adds latency and token cost.
Cumulative tokens the org can use per rolling window (ASSISTANT_TOKEN_BUDGET_WINDOW, default 24 hours) before new AI requests are refused. Blank uses the install default (ASSISTANT_TOKEN_BUDGET); 0 means unlimited.
Cost budget (USD)
A spend cap over the same window, enforced only for models with a known price. Blank means no cap.
Claw coordinator
Off by default. Turns on the Claw scan for this org: Claw watches your signals, routes each to an agent and queues a proposal. It also drives scheduled flows, and while it is on, AI review of change requests defaults to Auto on open.
Redacted fields
Extra SIEM and alert field names, on top of a built-in list (email, password, token and similar), redacted before data leaves the appliance for a model provider. People still see the original.
Budgets are a hard stop
Hitting the token or cost budget returns 429 for chat, agent runs, flow runs and previews until usage ages out of the rolling window; there is no soft-warning mode. Cost caps cannot see self-hosted and compatible models (they have no price entry), so a token budget is the only limit that reaches them.
Environment-only caps
The agent-run guardrails (AGENT_RUNS_DISABLED, the kill switch; AGENT_MAX_CONCURRENT, default 3; AGENT_MAX_RUNS_PER_DAY, default unlimited) and CLAW_ENABLED=false, which stops the Claw scan for every org, are environment variables, not settings on this page.
Usage
Talon Claw → Usage (/claw/usage, admins only) shows where the budget went over today, 7, 30 or 90 days: spend, model calls, tokens and the share of input served from the prompt cache; spend by day; and spend ranked by feature, by agent and by model, with the token and cost budget beside them when one is set. Every model call Studio makes is recorded under a feature: chat, agent runs, automation steps, stream classification, SIEM migration, Sigma conversion review, ATT&CK mapping, threat-intel extraction, synthetic events, memory upkeep, search intent, commit messages, connection tests, run reviews and agent improvement proposals. Each one counts toward the budgets above. Usage recorded before v1.0.2-beta reads Other (unlabelled).
The Usage tab here (admins only) covers the last 30 days: requests, input and output tokens, cache hit rate and an estimated cost, plus a per-model breakdown. Cost is computed when each request is recorded, from a price table in the code (per-million-token USD prices for the native Anthropic, OpenAI and xAI models), and is never recomputed afterwards.
"Est. cost (partial)"
Self-hosted and compatible models deliberately have no price entry, so their usage shows as unpriced rather than a fake $0; when some usage is unpriced the tile reads Est. cost (partial). Check prices against your own provider agreement.