Runic
AI SETTINGS

Runic’s Claude, or your own Anthropic, OpenAI or Google key

The owner decides whose key pays for AI; an account can bring its own key from any of the three providers and every AI feature runs on it, same prompts, same self-checks, no monthly budget.

Try it free →Read the docs

What it is

AI in Runic has always been proxied: the browser never talks to a model provider, the server does, and the platform key lives in the server environment. AI settings keeps that shape and adds a second source of keys. The owner opens the staff console, chooses the AI tab and picks one of three modes. Runic’s Claude runs everyone on the platform key and ignores stored keys. Users bring their own key turns AI off for an account until it stores a key. The third mode uses the platform key by default and switches an account to its own key the moment one is stored.

An account manages its key under Settings → AI (avatar menu → Settings; AI is the first section). A note states the mode in force — on Runic’s key: “AI features on this workspace run on Runic’s own Claude models. There is nothing to configure — chat, generation and autocomplete are ready to use.” — and when a key is allowed the form offers three providers, Anthropic, OpenAI and Google, an API key field, and two optional model fields: “Fast model (autocomplete, summaries)” and “Main model (chat, generation, edits)”, with each provider’s defaults as placeholders.

How a key is saved

Save sends the key once. The server calls the provider’s model list with it (15-second timeout): a refusal comes back as “OpenAI rejected that key — check it and try again.”, an outage as “Couldn’t reach OpenAI to check the key — try again in a moment.”, and nothing is stored in either case. If you named a model, it has to appear in that list; a name the key cannot see is refused with “‘x’ isn’t available on this OpenAI key.” Only then is the key encrypted with AES-256-GCM under a server-side secret and written with a seven-character hint such as sk-…7890. The response carries the hint, the provider and the model picks, never the key; the panel confirms “Saved — AI features on this account now run on OpenAI.” and shows a green summary — “Using your OpenAI key sk-…7890 · fast: gpt-5-mini · main: gpt-5” — with Change and Remove.

From then on every AI request for that account decrypts the key inside the request that spends it. Ten saves an hour, a key of at most 512 characters with no whitespace, model names of letters, digits and a few punctuation marks. If the server has no key-storage secret yet, the form is disabled with an amber note asking the owner to set AI_KEY_SECRET. The key is deleted with the account; the data export lists the provider, the hint and the model picks only.

Which model answers

Runic routes each action to a tier rather than a model. Light tier: autocomplete, archive summaries, and every planned document of a whole system except the main flow. Mid tier: chat, inline and selection edits, the system plan, the main flow, a single Create with AI document, the interview and build questions. Heavy tier: whole-document chat turns — the deep review, attachments, redesigns — and the build spec. On Runic’s key the tiers are Claude Haiku 4.5, Claude Sonnet 5 and Claude Sonnet 5 again; Opus 5 runs on the heavy tier only when the owner sets ANTHROPIC_MODEL_HEAVY. OpenAI’s defaults are gpt-5-mini and gpt-5, Google’s Gemini 2.5 Flash and Gemini 2.5 Pro. Your Fast model replaces the light tier; your Main model replaces both mid and heavy. The label under each chat reply names the model that actually answered.

Thinking is switched on for four tasks only — the system plan, the build questions, the build spec and the heavy chat turn — and off for everything else — stated explicitly on the models that would otherwise think, with low effort where the model takes an effort setting — because a document or an edit is checked by the deterministic parse-and-audit pass instead. Web research in the chat uses each provider’s own search tool. The system prompt and, for a whole system, the plan are sent as cached blocks; long replies stream; a 429 or a 503 is retried twice with a short backoff while the route still has time; a provider that rejects the thinking or effort fields is re-sent without them.

Budget and usage

On Runic’s key each account may spend an estimated $5 in a calendar month; past that, AI features answer “This month’s AI budget on Runic’s key is used up — add your own key under Settings → AI or wait for the 1st.” The budget never applies to an account on its own key, which its provider bills directly. The section ends with a usage line read from the cost ledger: on Runic’s key “This month: 12 calls · $0.41 of $5.00 on Runic’s key.” (red with “— used up until the 1st; add your own key to keep going” once exceeded; “(no limit on this account)” when exempt); on your own key “This month: 12 calls · ~$0.41 on your own key (list-price estimate).” Hourly limits per route apply on every key.

What does not change

The prompts are identical on every provider, and so is everything wrapped around them: the deterministic parse and structural audit after a generation, the single bounded repair pass that is kept only when it scores better and loses no content, the diff-and-apply flow for proposals, the checkpoints, the hourly limits. The only difference an account sees is the model label, the absence of the monthly budget, and the bill, which goes to its own provider account.

How to use it

  1. Owner picks the mode
    Staff console, AI tab: Runic’s Claude, users bring their own key, or Runic’s key unless an account stored its own. The change is audited.
  2. Open AI settings
    Avatar menu → Settings → AI. The section states the mode in force and shows the key form when a key is allowed.
  3. Pick a provider and paste the key
    Anthropic, OpenAI or Google. Leave the model fields empty for the defaults or name models the key can use.
  4. Save
    Runic checks the key against the provider, refuses a bad key or an unavailable model, then stores it encrypted and shows a masked hint.
  5. Use AI as before
    Chat, Create with AI, inline edits and autocomplete run on the stored key, with no monthly budget. Change or Remove from the same screen; the usage line at the bottom shows this month’s calls.

Commands & shortcuts

Runic’s ClaudePlatform key for everyone; stored keys are ignored
Users bring their own keyAI is off for an account until it stores a key
Runic’s Claude, own key overridesPlatform key by default, a stored key takes over
Fast modelAutocomplete, archive summaries, planned documents of a system
Main modelChat, edits, questions, the plan, generation, the deep review and the build spec
This month: N calls · $X of $5.00 on Runic’s key.the usage line; on your own key it reads “~$X on your own key (list-price estimate)”
AI_KEY_SECRET · ANTHROPIC_MODEL_HEAVYowner settings: the secret that encrypts stored keys; the model on the heavy tier

Tips

  • Leave the model fields empty unless you have a reason; the defaults match how Runic tunes each action.
  • A key that works in the provider’s console but is refused here usually lacks access to the model you named, so try saving without the model fields first.
  • Bringing your own key is the way past the $5 monthly line: the budget counts only calls Runic’s key paid for, and hourly limits are the only cap left.
  • Teams on a shared provider account can each store the same key; usage then shows up under one provider bill, with each account’s usage line estimating its own share.
  • If the owner switches the workspace to Runic’s Claude, your stored key stays saved but idle, and comes back into use if the mode changes again.

Questions

Can Runic staff see my key?

No. It is encrypted at rest under a server-side secret, decrypted only inside the request that uses it, and the interfaces (settings screen, admin console, data export) only ever carry a masked hint.

Which providers are supported?

Anthropic (Messages API), OpenAI (Responses API) and Google Gemini (generateContent). The same prompts run on all three.

Do my documents get stored by the provider?

Runic sends requests with storage disabled where the provider offers the option, and never sends more than the request needs. What each provider retains beyond that is governed by your agreement with them.

What happens when my key stops working?

AI replies say “Your API key was rejected by the provider — check it under Settings → AI.” Nothing else in the editor is affected. In the mode where Runic’s key is the default, a key that cannot be decrypted falls back to the platform key.

Does Opus 5 ever answer?

Only when the owner has set ANTHROPIC_MODEL_HEAVY, and only on the heavy tier. By default the heavy tier is Sonnet 5 with thinking on; at $5 in and $25 out per million tokens against Sonnet’s $3 and $15, the judged quality gap does not pay for Opus on these tasks.