Agent Harness

Interface language

Model (pick from the providers above)

Enabled Model (pick from the providers above)

Enabled Search provider Search API key only for tavily/brave/serper; stored in config.yaml Request timeout (s) Max search results Fetch text cap (chars) Proxy http(s) only; empty follows the environment's HTTP_PROXY/HTTPS_PROXY (on macOS also the system proxy settings)

Request timeout (s) Transport retries Max response tokens Max steps Tool calls per step Bash timeout (s) Compact at (% of context window) Consolidate after (messages) Keep recent (messages) Background workers Model-written summaries Cache the prompt prefix System role

Workspace root of the file tree and the agent's working directory Editor limit (MB) larger text files offer a download

Current password New password

Chat

Start a conversation

Settings

Language

LLM providers every model below appears in the chat model picker as <provider>/<model>

Fallback model answers when a request to the active model fails; pick one on a different endpoint

Media sidecar describes image attachments so a text-only main model can use them

Web tools web_search + web_fetch; duckduckgo, bing and baidu are free and need no key

Agent parameters defaults are the HarnessConfig defaults

Web

Security