onyx

Facts

Openness

Public code, licence not stated

Language

Python

Pinned version

23be912 (2026-10-01)

Topics checked

28 of 33

Kind of agent

General framework

Source repository (opens another site)

What it does notably well

• Budgets

  When deep research is used, onyx checks elapsed time at the start of each orchestrator cycle (30 min, DEEP_RESEARCH_FORCE_REPORT_SECONDS) and then writes the final report from…

  onyx fits chat history to 95% of the model's max_input_tokens (GEN_AI_INPUT_TOKEN_SAFETY_MARGIN, default 0.05); it then sends max_tokens = min(model max output, context window…

• Compaction

  onyx sizes the verbatim tail kept after compression as RECENT_MESSAGES_RATIO = 0.2 of the current history tokens (not of the context window), converts it to a token budget at run…

  onyx documents its layered context design in-tree: chat/README.md has a "Reasons / Experiments" section giving empirical reasons for the placement of custom-agent instructions…

  onyx keeps tool-call and result pairing at two sites: run_llm_loop appends one assistant message carrying all calls of a cycle followed by their tool-response messages in one…

• Concurrency

  onyx uses one pool worker per selected model, one thread per tool call, and a per-call isolated litellm HTTPHandler for true OpenAI models and for the Anthropic, Bedrock and…