Models, providers & secrets

This page covers what powers an agent (models and gateways) and the two credential stores behind it.

Models & gateways#

Each agent has a model and a runtime gateway. Gateways: OpenClaw (the default), Anthropic, Claude Code, Ollama, and OpenRouter.

Note: Default agents route through OpenClaw even when their model name carries an openrouter/… prefix, OpenClaw resolves the downstream provider internally. The non-OpenClaw gateways are for agents explicitly routed to them.

Provider API keys, where they go#

Set provider keys (Anthropic, OpenAI, Google, OpenRouter, Z.ai) in Settings → AI & Models → Model Providers.

Warning: Provider keys must go in Model Providers, not the generic Secrets tab. A key saved in Secrets is stored but not wired into the runtime, the agent will keep using the old key and you'll think it's set when it isn't. Same encrypted storage either way; only the Model Providers save routes it.

One key is used even if you never run an agent#

Every provider here serves agents you run, with one exception. The plain-English summary on approval cards (what an action does and why it's risky) is generated with your OpenRouter key, called directly rather than through the gateway. It needs no license and no runtime, so it works on a governance-only install where the agent builder is locked.

Without an OpenRouter key, approval cards still work; they just arrive without that summary. Nothing else changes, and no other provider substitutes for it.

ChatGPT / Codex subscription#

Instead of a pay-per-token OpenAI API key, you can connect a real ChatGPT Plus/Pro or Codex subscription. In Settings → AI & Models → Model Providers, use Connect on the ChatGPT / Codex row and complete the sign-in. The runtime handles the OAuth flow and refreshes the token itself; damn.dev only records that it's connected, never the token. Agents then run on openai-codex/gpt-5.5, and connecting sets that as the default model for new agents in the workspace.

Tip: Connecting is additive. Your other providers stay available, existing agents keep their current model until you choose Apply to all agents (or change them one at a time), and the subscription falls back to your previous model automatically if it hits its usage cap.

Respects OpenAI's terms: a ChatGPT subscription is licensed for personal use by the subscriber, so damn.dev only lets a connected subscription run agents in a single-user workspace. If a second person joins, agents automatically move off the subscription back to their previous model. To run a team on GPT, connect a company OpenAI API key (billed to the organization) in the same Model Providers panel, it carries no such limit.

Secrets vs. Environment#

Two stores, on purpose:

• Secrets: credentials skills use (API tokens, passwords). Encrypted with AES-256-GCM under ~/.damn-dev/secrets.key. Agents reference them as ${KEY}; the value is substituted server-side and never shown to the agent.
• Environment: plaintext infrastructure config (paths, infra tokens) read at boot. Plaintext by design, encrypting it with a key sitting in the same directory would add nothing against a host-access threat.

Warning: Losing ~/.damn-dev/secrets.key means every secret is permanently unrecoverable. It's the one file with no other backup path. Both stores offer a Download button (the key, and the .env), use them, and keep the key somewhere off the host. This is your responsibility as the deployer.

Next#

• Per-agent access, granting a secret to specific agents only.
• Architecture, gateways & runtime, how the gateways fit together.