Omniscio documentation
Browse all documentation
  1. Getting Started19
  2. Sessions & Agents140
  3. Inbox & Notifications72
  4. Projects & Tasks102
  5. Automation & Scheduling77
  6. Knowledge & Memory29
  7. AI Features74
  8. Integrations115
  9. Plugins & Marketplace44
  10. Cloud & Teams53
  11. Settings & Customization81
  12. Account & Billing29
  13. Troubleshooting84
  14. CLI & API Reference28
  15. Legal & Policies5
  16. Uncategorised20

Refuseless

Refuseless hosts a named lineup of refusal-removed ("abliterated") open-weight models behind an OpenAI-compatible API. In Omniscio it runs as a proxy-backed provider: the Claude Code harness is pointed at Omniscio's local model proxy, which forwards the conversation to Refuseless on your own API key. Nothing about your sessions is pooled — you pay Refuseless per token.

Use it when you want a strong open-weight coder without the refusal behaviour the stock weights ship with: the vendor's own benchmark framing is frontier coding and security work.

Setup

  1. Get a key. Create one at refuseless.com. Keys start with abl_.
  2. Paste it into Settings → the Refuseless key field (refuselessApiKey).
  3. Turn on Refuseless sessions with the Allow Refuseless sessions toggle (allowRefuselessSessionSpawn). It ships off: no Refuseless session can start until you enable it. Both the key field and the toggle are findable through Settings search by typing "refuseless".

Nothing else to install — the provider runs on the same Claude Code binary every proxy-backed vendor uses.

Models

Model Context Price per 1M tokens (in / out)
glm-5.3-abliterated (default) 1M $6 / $9
glm-5.3-flash-abliterated 1M $1.25 / $4

Prices are the vendor's published rates and drive Omniscio's spend estimates. Refuseless also serves three further models in some accounts — deepseek-v4.1-flash-abliterated, deepseek-v4.1-flash-fast and minimax-m3.1-flash — which are deliberately not in the picker until the vendor publishes a rate for them: a catalog entry with no real price would book their spend as zero.

The provider streams reasoning on its own channel, so a thinking model's scratchpad renders separately from its answer.

How it works

  • Sessions launch exactly like every other proxy-backed vendor: the Claude binary runs, with its base URL and auth pointed at the local model proxy.
  • The proxy's openai-compatibility upstream for Refuseless is https://api.refuseless.com/v1 (overridable with AMC_REFUSELESS_BASE_URL).
  • Model ids are exposed to the picker as refuseless/<upstream-id>; the prefix is stripped before the vendor call.
  • Cost is reported as estimated — token counts come from the harness, prices from the catalog above.

Code references

  • src/shared/providers/proxy-provider-catalog.ts — the one catalog entry everything derives from.
  • src/shared/types/settings/ai-features-settings.ts — refuselessApiKey (encrypted at rest, stripped from renderer, CLI and backup reads).
  • src/shared/types/settings/accounts-providers-settings.ts — allowRefuselessSessionSpawn.
  • src/shared/providers/provider-lifecycle-coverage.ts — its lifecycle column (mirrors the other catalog resellers).
  • src/main/db/migrations/ — the migration widening the pending-spawn provider allow-list.

See also