Omniscio documentation
Browse all documentation
  1. Getting Started19
  2. Sessions & Agents134
  3. Inbox & Notifications68
  4. Projects & Tasks98
  5. Automation & Scheduling79
  6. Knowledge & Memory28
  7. AI Features71
  8. Integrations107
  9. Plugins & Marketplace36
  10. Cloud & Teams53
  11. Settings & Customization71
  12. Account & Billing28
  13. Troubleshooting79
  14. CLI & API Reference27
  15. Legal & Policies5
  16. Uncategorised17

Venice Provider

Venice AI as a session provider: refusal-stripped (uncensored) models — GLM 4.7 Flash Heretic and Abliterated Large V2 — run inside the Claude Code harness on your own Venice key, with Venice's cached-input discount recorded correctly. How to switch it on, what each model costs, and the output cap and context limits to know.

Venice AI serves refusal-stripped ("uncensored") models over an OpenAI-compatible API. Omniscio runs them inside the normal Claude Code harness, so a Venice session gets the same tools, files and approvals as any other session. You pay Venice directly on your own key.

The models

Model Input Cached input Output (per 1M tokens) Context
GLM 4.7 Flash Heretic (default) $0.07 $0.035 $0.40 200K
Abliterated Large V2 $3.00 $0.30 $5.00 1M
  • Cached input is what makes Venice cheap for agents. An agent re-sends the same long instructions on every turn. Venice bills the repeated part at the cached rate, and Omniscio records it at that rate too. Measured on 2026-10-13: a warm turn on each model came back with over 55,000 tokens read from cache, and the recorded cost matched Venice's own bill.
  • Venice is a pickable provider, not a hidden reseller, because no other vendor serves these models. Pick it like any other engine in the session's Provider menu.

Switching it on

  1. Add your Venice API key: Settings → Accounts → the Venice card. Venice creates keys at venice.ai/settings/api; use an inference key.
  2. Turn on Allow Venice sessions on the same card. It is off by default, and agents cannot start Venice sessions until it is on.
  3. Start a session and choose Venice as the provider.

The key is stored encrypted, stays in the background process, and never reaches the interface.

Limits to know

  • Replies are capped at 24,000 tokens. Venice refuses the harness's usual 32,000-token output request on GLM 4.7 Flash Heretic, so every Venice session asks for 24,000 at most. The cap covers the whole lane because switching models mid-session keeps the settings the session started with.
  • Models with a 128K window are not offered (for example Venice Uncensored 1.2). The harness compacts a conversation near 167K tokens, so a 128K model would overflow first.
  • Venice's own requests go through the local model proxy. Venice publishes no Anthropic-compatible endpoint. The proxy carries Venice's cache counts through correctly, which was checked end to end with the real claude binary.
  • Prompts, files and tool results go to Venice when you pick it. It is listed in the subprocessors inventory (row 79).

Related

Last verified 2026-10-14