Configuration file example
Parent: Configuration.
[model]
provider = "openai"
model = "gpt-5.6-sol"
auth = "api-key" # or "none", "codex", "anthropic-api-key", "google-api-key", "github-copilot", "xai-api-key", "xai-oauth", "moonshot-api-key", "ollama-api-key", "ollama-cloud-api-key", "ollama-cloud-device", "poolside-api-key", "openrouter-api-key", "openrouter-oauth", "kimi-oauth", "qwen-token-plan-api-key", "meta-api-key", "meta-muse", "minimax-api-key", "mistral-api-key", "opencode-go-api-key", or "opencode-api-key"
reasoning = "medium" # off, minimal, low, medium, high, xhigh, or max
fast_mode = false # faster serving for models that support /fast; bills at a higher rate
favorite_models = []
[model.aliases]
# deep = "anthropic/claude-opus-4-8"
# fast = "gpt-5.6-luna"
[output]
max_output_bytes = 64000
[compaction]
auto_compact = true # set false to opt out
compact_threshold_percent = 85
compact_target_percent = 50
[internal_agents.session-title]
# provider = "openai"
# model = "gpt-5.6-sol"
# auth = "api-key"
[internal_agents.advisor]
# provider = "anthropic"
# model = "claude-fable-5"
# auth = "anthropic-api-key"
# reasoning = "high"
[internal_agents.compaction]
# provider = "openai"
# model = "gpt-5.6-luna"
# auth = "api-key"
# reasoning = "low"
[internal_agents.permission-classifier]
# provider = "openai"
# model = "gpt-5.6-luna"
# auth = "api-key"
# reasoning = "low"
[web_search]
mode = "auto" # auto, backend, or off
backend = "openai" # openai, exa, brave, or firecrawl
[web_search.openai]
connection = "api" # api or codex (fixed ChatGPT endpoint)
# api_base_url = "https://api.openai.com/v1"
[web_search.exa]
connection = "api" # api or mcp; missing API credentials do not switch transports
# api_base_url = "https://api.exa.ai"
# mcp_url = "https://mcp.exa.ai/mcp"
[web_search.brave]
# api_base_url = "https://api.search.brave.com"
[web_search.firecrawl]
# api_base_url = "http://localhost:3002" # omit for Firecrawl Cloud
# Set keys in the environment (FIRECRAWL_API_KEY, EXA_API_KEY, BRAVE_SEARCH_API_KEY), not this file.
# MCP stays inert while this table has no enabled server entries.
# See /integrations/mcp for stdio and Streamable HTTP examples.
[mcp.servers]
# Read-only REST tools from an OpenAPI spec. See /integrations/openapi.
# [openapi.github]
# spec = "https://raw.githubusercontent.com/github/rest-api-description/main/descriptions/api.github.com/api.github.com.json"
# base_url = "https://api.github.com"
# auth = { type = "bearer", token_env = "GITHUB_TOKEN" }
[providers.ollama]
base_url = "http://127.0.0.1:11434/v1"
[providers.custom.composer]
base_url = "http://127.0.0.1:8787/v1"
# catalog = "llmgateway" # optional models.dev slug for context, price, and reasoning
# catalog_mode = "model-id" # look up unsplit slug/model ids in models.dev
# edit_tool = "apply_patch" # auto preference: str_replace or apply_patch
# request_compression = "zstd" # or "gzip"; only if the host decodes that request encoding
[codemode]
mode = "on" # "on": natives and codemode declared; "only": tools hidden unless tool_search loads them
[behavior]
advisor_mode = false
enable_subagents = true
agent_concurrency = 10
workspace_rewind = true # native file-tool checkpoints; restart after changing; false opts out
edit_tool = "auto" # auto, str_replace, or apply_patch
permission_mode = "bypass" # bypass, auto, allow_edits, plan, or supervised
rtk = true
inline_shell = "bash" # bash default on macOS/Linux; powershell on Windows
# credential_store = "os" # or "file"; omit until first /login chooses
[prompt_templates]
review = "Review this code for correctness, security, and maintainability."
"explain-tests" = "Explain how these tests cover the expected behavior."That is ~/.rho/config.toml. Every setting sits in a table grouped by purpose; a key at the top level is a load error, except in a 2.x file, which Rho rewrites at startup.
The interactive TUI keeps its own settings at the top level of ~/.rho/clients/tui.toml:
show_reasoning_output = true
zen_mode = false
output_streaming = "live" # live, paragraph, off; /config → Display → Output streaming
max_tool_output_lines = 10
prompt_history_limit = 1000 # 0 disables saving; at most 10000
cache_miss_notices = false
show_header_hints = true # keyboard hints under the session header; sign-in hints stay while signed out
notifications = true # OSC 9 or bell when a turn finishes or needs you while unfocused; off under Herdr
editing_mode = "default" # or "vim" for vim-style normal/insert editing in the composer
check_for_updates = trueThe theme and keybinding chords live in ~/.rho/tui.toml:
theme = "terminal" # terminal, a built-in id from /theme, or a custom ~/.rho/themes/<id>.json stem
[keybindings]
search_prompt_history = "ctrl+r"
search_transcript = "ctrl+f"
# reset_conversation = "ctrl+shift+n" # same as /new; unbound by default
open_editor = "ctrl+g"
jump_to_bottom = "ctrl+end"
toggle_tool_output = "ctrl+o"
cycle_streaming_mode = "alt+s"
cycle_permission_mode = "alt+m"
insert_newline = "ctrl+j"
queue_prompt = "alt+enter" # ctrl+enter always works as a fallback
paste_image = "ctrl+v"
edit_pending_input = "alt+up"
manage_pending_input = "alt+q"
cycle_pinned_model = "ctrl+p"
cycle_pinned_model_back = "ctrl+shift+p"
undo = "ctrl+z"
redo = "ctrl+shift+z" # needs a terminal that reports ctrl+shiftKeybindings use +-separated modifiers and keys. Supported modifiers are ctrl, alt, and shift; supported named keys include enter, esc, tab, arrow keys, home, end, pageup, pagedown, backspace, and delete. Single-character keys can be used directly. queue_prompt queues a follow-up during a model turn (default alt+enter); ctrl+enter is always accepted as a fallback because some terminals bind alt+enter to fullscreen. undo and redo act on the composer and the side chat prompt. editing_mode = "vim" in clients/tui.toml turns on vim mode for the main composer. Keybinding changes take effect when Rho starts.
The full saved file can also include model overrides for reserved internal agents. Each entry under [internal_agents] selects the provider, model, and auth used by that role. An internal agent with no entry follows the active conversation selection. [providers.ollama].base_url and [providers.custom.<name>].base_url set OpenAI-compatible endpoints used for those hosts' chat, model refresh, and health checks. First-run setup does not write [providers.ollama]; /login ollama stores the API base and an optional key. [providers.custom.<name>].catalog optionally borrows a models.dev provider for context, price, and reasoning. Web search API keys are normally stored in the configured credential store rather than config.
Ollama's provider-specific API base uses its own section and does not affect other providers. It appears after /login ollama or a hand edit:
[providers.ollama]
base_url = "http://127.0.0.1:11434/v1"Custom OpenAI-compatible hosts use a name you choose. They speak Chat Completions unless you pick Custom · Responses in /login, or set api = "responses". catalog is optional and borrows a models.dev provider for context, price, and reasoning. catalog_mode = "model-id" looks up the unsplit slug/model id instead and cannot be combined with catalog:
[providers.custom.composer]
base_url = "http://127.0.0.1:8787/v1"
[providers.custom.cliproxyapi]
base_url = "http://127.0.0.1:8317/v1"
catalog = "llmgateway"
[providers.custom.litellm]
base_url = "http://127.0.0.1:4000/v1"
api = "responses"See Ollama and Custom OpenAI-compatible hosts for setup, model refresh, and endpoint limits.