Loading...
RightNow Agent is a local tool runner with remote model inference. Its launchers prepare configuration, the selected provider produces tool calls, and approved tools execute on the user's machine.
launcher or direct binary
|
+--> seed or synchronize user config
|
+--> interactive session or one-shot prompt
|
+--> OpenAI-compatible provider request
|
+--> validate model tool call
|
+--> permission decision
|
+--> local file, search, web, or terminal tool
|
+--> result returns to the model looprightnow.ps1 uses the PowerShell migration and sync-stamp helpers. rightnow.sh invokes the binary's early config sync dispatcher. Both use config/rightnow.toml and default the user home to ~/.rightnow.
Two commands dispatch before the main parser: config sync --config <path> --template <path> and telemetry speed [--log <path>] [--since <when>] [--json]. Other arguments continue to the main binary parser.
The shipped model entry calls RunInfra at https://api.runinfra.ai/v1, reads RIGHTNOW_API_KEY, and selects qwen3-8-flash-next. User-defined provider entries use the same parser, and -m <model-entry> selects one for an invocation. No local inference process, model-weight download, GPU scheduler, or local embedding model is shipped.
RightNow assembles the system prompt, configured rules, workspace context, and selected tool schemas before each provider request. The launchers do not inject a system prompt override. --rules appends a <human_rules> block.
File and terminal tools run locally under the selected permission policy. Web tools make network requests when called. Configured MCP servers and hooks keep their own command, credential, and network behavior.
The shipped template enables the warm daemon. It reuses the lightweight client process and parsed runtime state, not model weights. A connection timeout can fall back to the embedded process path before a turn starts.
Local JSONL diagnostics are written to ~/.rightnow/logs/unified.jsonl. Named events include:
shell.turn.inference_done
shell.turn.loop_phase_timings
shell.handle_prompt.done
shell.tool.exec_done
shell.turn.speed_shadowThe speed-shadow record includes model-call and tool-round counts, narration-only and narration-with-tool counts, narration size and estimated savings, one-call transaction outcomes, and tool-profile fields.