Skip to main content
Voss keeps harness configuration narrow in v0.1. The most important setting is the preferred model.

​
Show or edit config

The default config path is:

​
Recall and context options

Code recall injection can be disabled when you want the harness to skip the automatic recall section:
Context packing can also be disabled globally, or per run with voss do --no-pack:

​
Switch model and auth in a session

Inside the Go terminal client:
/model (or /models) searches available models and /auth chooses Claude, Codex, or API authentication. Subscription choices reuse your local Claude or Codex login; API choices need a configured key. Successful switches persist across launches and saved-session resumes. voss ui --auth codex overrides auth for one launch.

​
Login flow

Run the sign-in wizard, or use the provider CLIs directly and tell Voss which auth source to prefer:
Under --auth=auto, Voss checks keychain keys stored by voss login, then ANTHROPIC_API_KEY / OPENAI_API_KEY, then Claude Code OAuth, then Codex auth. Inside voss chat, /login shows local auth status and refreshes existing credentials when possible.

​
Network tools

Network tools (web_fetch, web_search, and MCP) are off unless you pass --allow-net or set:

​
Jev judgments

Jev judgments are off by default and opt-in per project in .voss/config.yml:
A Jev key is read from TYPESAFE_API_KEY or the keychain. VOSS_JUDGMENTS=off stops every Jev call. voss judge activate code_recall switches recall reranking to active after checking the published evaluation gates. Model, timeout, and per-turn call and dollar limits live under [judgments] in config.toml.

​
Provider safety

Session persistence must not store provider credentials. Credentials should remain in provider-specific auth storage or environment configuration.
Run voss doctor after changing provider or model settings.