AI providers

Every agent runs its turns with one AI provider. You add the keys once under Settings → AI providers, then pick a provider, a model and an effort for each agent in the agent editor (Team, then the pencil on an agent). You can mix them: a Claude lead, a Codex engineer and a Gemini reviewer on the same project is fine.

Keys are encrypted on your server (AES-256-GCM), and the panel can replace or remove a key but never shows it again. During a turn the key is handed to the provider's command-line tool inside that turn's container, so an agent could read it there. Use a key made for Klatch, with a spending limit set at the provider, and replace it if you ever doubt it. (Keys for connections are different: those stay in the hub and agents never see them.)

You pay the provider directly, under your own account. Klatch adds nothing on top.

Claude

Agents run Claude Code. Models: Fable (newest and most capable), Opus (deep work), Sonnet (fast and capable), Haiku (quick and cheap). Effort goes from Low to Max.

API key (recommended)

Create a key at console.anthropic.com → API keys, paste it into the Claude card and press Check. This is the sign-in Anthropic asks products like Klatch to use.

Subscription token

A token from claude setup-token uses a personal Claude plan. It is meant for your own use: Anthropic's terms for software that other people use call for API keys, so read them before you choose this.

When Claude itself reports its weekly limit, Klatch slows down at 70% (one turn at a time, only work you started) and pauses at 95%, then resumes when the limit resets. You can switch this brake off in Settings.

Other services that speak Claude's API

Choose Other service on the Claude card to send Claude Code to a different endpoint. Presets:

Each preset maps Fable, Opus, Sonnet and Haiku to that service's models; you can change the mapping. For OpenRouter, type the model ids you want (for example qwen/qwen3.7-flash).

On these services Claude Code cannot price the turns correctly, so Klatch shows tokens but no cost, and the weekly-limit brake does not apply.

Codex (OpenAI)

Agents run OpenAI's Codex CLI. Create an API key at platform.openai.com → API keys and paste it into the Codex card. Models include GPT-6.1 Sol (most capable), GPT-6.1 Terra (balanced) and GPT-6 Luna (fast and cheap). Codex reports tokens, not cost, so Klatch shows tokens.

Antigravity (Gemini)

Agents run Google's Antigravity CLI, which replaced Gemini CLI in June 2026. Create a key at aistudio.google.com → API keys and paste it into the Antigravity card. Models include Gemini 4 Argon (most capable) and Gemini 3.8 Flash (fast).

Checking a key

Check on each card makes one small request and shows the provider's own answer if it fails (a wrong key, no credit, a model your account cannot use). A subscription token can only be checked by running a turn, so it is checked the first time an agent uses it.

When a provider fails

A failed turn says why in the chat, in the provider's own words, and the agent's next message starts a fresh turn. Retry on a message runs it again. A provider that is down does not stop agents that use a different one.

Costs

Every agent message shows what its turn cost (when the provider reports it) and how long it took. Usage totals them per agent and per project.