Narratium custom API setup: base URL, key, model.

Quick answer: the values Narratium needs.

Narratium's LLM connection is the standard OpenAI-compatible triple, entered in its API settings: a base URL, an API key, and a model name. For this gateway those are https://api.apisrouter.com/v1, your sk-... key, and any catalog id, deepseek-v4-flash being the sensible first pick for story traffic. The project's README recommends exactly this class of connection: OpenAI-format APIs directly, a unified multi-model endpoint for variety, or Ollama and LM Studio for local models, all through the same configuration surface. Exact field labels vary between the web build and versions, so treat labels loosely and values exactly: URL ends at /v1, key pastes clean, model id matches the catalog string character for character.

Base URL:  https://api.apisrouter.com/v1
API key:   sk-...              (from APIsRouter)
Model:     deepseek-v4-flash   (typed exactly; any catalog id works)

What Narratium is, and its status, stated plainly.

Narratium (Narratium/Narratium.ai on GitHub) pitched itself as VSCode for roleplay: an open-source client for AI-driven storytelling, worldbuilding, and long-form interactive fiction, with SillyTavern character card compatibility, memory persistence, and a plugin system. The code is AGPL-licensed, it attracted a real community quickly, and it remains one of the more interesting open clients in the scene. Now the part most setup guides omit: the GitHub repository is archived and read-only (GitHub shows the archive date as March 30, 2026). Archived does not mean broken, the code runs exactly as well as the day it was frozen, but it means no maintainer is shipping fixes, dependency updates, or security patches, and the hosted instances associated with the project should be treated as best-effort at most. If you adopt Narratium today, the dependable route is running it yourself from the source or a release, and the realistic expectation is a frozen feature set. That status changes the API advice in one specific way: a frozen client is exactly the situation where you want the model side to stay flexible. Narratium cannot add new provider integrations anymore, but an OpenAI-compatible endpoint it already speaks keeps serving whatever new models appear behind that endpoint. The gateway pattern effectively future-proofs a client that is no longer moving.

Setup walkthrough.

Troubleshooting follows the universal pattern: prove the three values with a curl request first, and if curl succeeds while the app fails, the usual culprits are a URL with a stray path segment, whitespace in the key, or a browser CORS block on web builds, visible in the developer console. Since no maintainer will ship a fix for app-side quirks anymore, the curl-first habit matters a little extra here: it tells you immediately which side of the connection you can actually do something about.

• Get Narratium running: self-host from the GitHub source or a release (the repo is read-only but fully cloneable), or use a hosted instance if one you trust still operates.
• Open the settings and find the API / LLM configuration section.
• Select the OpenAI-format connection type rather than Ollama, since you are pointing at a hosted endpoint.
• Enter base URL https://api.apisrouter.com/v1 and paste your key.
• Type the model id exactly: deepseek-v4-flash to start.
• Save and send one short message in a fresh story to confirm the round trip.

Cards, worlds, and what survives if you leave.

Narratium reads SillyTavern-compatible character cards, which is the single most important compatibility fact about it. Your characters are portable artifacts: cards written elsewhere import, and the work you put into personas and world notes is not locked to a client whose development has stopped. If you ever migrate, SillyTavern, RisuAI, and Wyvern all speak the same card lineage, and the setup pages linked below cover each. The same portability logic applies to the model side. Because Narratium talks to a generic /v1 endpoint, nothing about your provider setup is Narratium-specific: the same key and base URL drop into any other frontend unchanged. Between portable cards and a portable endpoint, the switching cost of the whole stack rounds to zero, which is the correct posture toward any archived project you still enjoy using.

Model picks for Narratium's long-form style.

Narratium's design centers long branching narratives, which weights model choice toward consistency over long horizons. deepseek-v4-flash carries everyday sessions at the catalog's lowest rates with dialogue the roleplay community consistently rates above its price. glm-5 and kimi-k2.6 rotate in for voice variety, and glm-5.2 earns detail-heavy chapters where its long-horizon instruction adherence outweighs its reasoning latency. claude-sonnet-4-6 remains the safe-for-work prose ceiling for chapters that deserve it. Because the frontend resends story context each turn, input tokens dominate cost, and a deliberate working window beats a maximal one; the context-length guide linked below does the math. Content policy is unchanged by any of this: model policies and applicable platform terms apply as written, and the policies-compared page covers which families fit which content ratings, factually.

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Should you still adopt an archived client?

An honest cost-benefit, since this page is the setup guide either way. In favor: the code is open, the interface is genuinely good for structured long-form stories, cards are portable in and out, and a generic endpoint keeps the model side current forever. Against: no security patches, no bug fixes, dependencies aging in place, and any hosted instance living on borrowed time. For a hobby client running locally against a metered key, that risk profile is acceptable to many people; for anything with other users or sensitive data, it is not. The middle path is the one this page has been describing throughout: run it yourself, keep your investment in the portable layers (cards, lorebooks, the endpoint), and treat the client as replaceable. Frontends in this scene come and go; a story backed by portable cards and a stable /v1 endpoint outlives all of them.

FAQ