user configuration

Configuration reference

Steiner is configured with YAML files, a small set of environment variables, and command-line flags. This page is the reference for every config field, its type, default and constraints, plus file locations, precedence, environment variables, expansion syntax and project trust.

For provider and model recipes (aliases, profiles, selection), see Provider and model setup. Command-line flags are listed in CLI.

File locations and precedence

Sources are applied in this order; later sources win:

1. Compiled defaults.
2. Global config: ~/.config/steiner/config.yaml (under $HOME).
3. Project config: <project>/.steiner/config.yaml, applied only once the project is trusted (see Project trust).
4. Environment variables.
5. Command-line flags that override config: --profile, --model, --verbose (sets logging.level to debug) and --unsafe (sets sandbox.enabled to false).

Both files are optional and neither is created for you. --config <path> replaces the project config path (the global file is still read); unlike the default path, that file must exist. With no providers, models or profiles configured, starting a session fails with no model configured; inspection commands such as steiner config still work.

Maps merge by key across files (providers, model definitions, profiles, tools, MCP and LSP servers, limits.tool_timeouts, and the string maps inside them). Scalars and lists are replaced by the later file.

Unknown keys are rejected. Durations are Go duration strings such as 250ms, 30s, 5m or 1h. A leading ~/ is expanded to your home directory in these path fields: logging.file, logging.compaction_log_file, diagnostics.dir, lsp.cache_dir, paths.writable_paths, paths.blocked_paths, paths.exclude_paths, sandbox.host_mounts.<index>.path, tools.<name>.exec, project_context.extra_files and project_context.ignore_files.

steiner config prints the resolved configuration. It includes secrets in the output, so redact before sharing.

Environment variables

Only the variables below override configuration; there is no generic STEINER_ prefix mapping. Integer variables that do not parse make config loading fail. Steiner reads a few other variables outside config loading: STEINER_GITHUB_TOKEN for update checks (see Installation), STEINER_USAGE_TELEMETRY and STEINER_USAGE_TELEMETRY_RUN for usage telemetry (see Usage and cache statistics), the variable named by providers.<name>.api_key_env, and any variable referenced by ${VAR} expansion below.

Variable	Effect
STEINER_MODEL	Active orchestrator model: a models.definitions alias or provider/model-id. An invalid value is silently ignored. --model wins over it; an invalid --model is also silently ignored.
STEINER_SUB_AGENTS_MAX_PARALLEL	Sets sub_agent.max_parallel.
STEINER_TUI_FPS	Sets tui.fps.
STEINER_MAX_TURNS	Sets limits.max_turns.
STEINER_MAX_TOKENS	Sets limits.max_tokens.
STEINER_TOOL_OUTPUT_MAX_BYTES	Sets limits.tool_output_max_bytes.
STEINER_MAX_PARALLEL_TOOLS	Sets limits.max_parallel_tools.
STEINER_LOG_LEVEL	Sets logging.level.
STEINER_LOG_FILE	Sets logging.file.
STEINER_COMPACTION_LOG_FILE	Sets logging.compaction_log_file.
KAGI_API_KEY	Fills search.kagi_api_key when it is empty.
BRAVE_API_KEY	Fills search.brave_api_key when it is empty.
STEINER_TRUST_PROJECT_CONFIG	1 or true trusts the current project's config for this run without prompting. Nothing is persisted. See Project trust.

Environment variable expansion in config values

Scalar values in config files can reference environment variables. Keys and comments are never expanded.

Form	Result
${VAR}	Value of VAR. Loading fails if VAR is unset. A set but empty VAR expands to an empty string.
${VAR:-default}	Value of VAR if set and non-empty, otherwise default (itself expanded). ${VAR:-} allows an empty value.
$VAR	Same as ${VAR}; has no default syntax.
$$	A literal $.

Backslash is not an escape character. An unquoted expansion is re-typed after substitution, so max_attempts: ${RETRIES:-5} and enabled: ${FLAG:-true} work; quoted values stay strings.

An undefined variable aborts loading at the first file that references one. The error lists every undefined variable in that file with its line, YAML path and name.

providers:
  openai:
    type: openai
    base_url: ${OPENAI_BASE_URL:-https://api.openai.com/v1}
    api_key: ${OPENAI_API_KEY}

Project trust

The first time Steiner runs in a directory it has not seen, it asks whether to trust that project before applying .steiner/config.yaml. The prompt appears whether or not the project has a config file. The trust key is the symlink-resolved absolute working directory, not a discovered git root, so a subdirectory of a trusted project prompts again.

When the project has a config, the dialog lists each leaf setting it overrides, compared against your global config (or the compiled defaults where the global file does not set it). Security-relevant fields are flagged with a leading !: sandbox, permissions, paths.writable_paths, paths.blocked_paths, paths.project_root_only, mcp.servers.<name>.transport, command, args, env, url, headers, approval and trust_annotations, lsp.servers.<name>.command, args and env, tools.<name>.exec and subcommand, and providers.<name>.base_url, headers, api_key and api_key_env. Existing values of api_key fields and anything under headers or env are shown as (set). Project values are shown unexpanded, so ${VAR} shows which variable would be used. If the project YAML fails to parse, the dialog shows the parse error instead of the diff.

Choices:

• a Always trust: recorded in ~/.config/steiner/trusted_projects.json; later config changes do not prompt again.
• s This session only: nothing is written.
• d Deny and exit: the default; Esc and Ctrl-C choose it. The process exits non-zero.

There is no revoke command: edit or delete trusted_projects.json. A corrupt or unrecognised-version file is deleted with a one-time notice; a read or permission error is fatal and leaves the file untouched.

The dialog needs an interactive terminal. With --exec or without a TTY, an untrusted project makes Steiner exit non-zero unless you pass --trust-project-config or set STEINER_TRUST_PROJECT_CONFIG=1 (or true). Both trust the project for that run only. steiner oneshot follows the same rule.

steiner config never prompts: it prints the resolved config including the project layer, prefixed with # project <root> is NOT trusted when the project is not in the trust store.

--unsafe and project trust are independent of each other.

Top-level keys

Path	Type	Default	Description
providers	map	empty	Model API connections. See providers block.
models	block	see section	Model definitions, discovery and profiles. See models block.
limits	block	see section	Turn, token, timeout and concurrency limits. See limits block.
sandbox	block	enabled	Bubblewrap sandbox. See sandbox block.
permissions	block	see section	Host capabilities granted to the sandbox. See permissions block.
sub_agent	block	enabled	Delegated child agents. See sub_agent block.
advisor	block	disabled	Stronger-model advisor. See advisor block.
oneshot	block	see section	Oneshot closeout. See oneshot block.
desktop_notifications	block	disabled	Desktop notifications. See desktop_notifications block.
update_check	block	enabled	Startup update check. See update_check block.
tools	map	empty	Custom subprocess tools. See tools block.
project_context	block	see section	Extra files added to the prompt. See project_context block.
paths	block	see section	Filesystem access limits. See paths block.
logging	block	see section	Session log files. See logging block.
diagnostics	block	disabled	Structured diagnostics capture. See diagnostics block.
context_management	block	see section	Context management baseline. See context_management block.
search	block	unset	Web search backend. See search block.
mcp	block	enabled	MCP servers. See mcp block.
lsp	block	disabled	Language servers. See lsp block.
modes	block	see section	Execution mode default. See modes block.
tui	block	see section	Interactive UI settings. See tui block.

TUI appearance preferences (accent, sidebar, backgrounds) are not config keys; they live in ~/.config/steiner/prefs.yaml. See Interactive mode.

providers block

A map of named provider connections. The key is an arbitrary name that models.definitions.<alias>.provider and provider/model-id references use. Setup recipes are in Provider and model setup.

Path	Type	Default	Description
providers.<name>.type	string	none	Required. One of openai_compat, ollama, lmstudio, openrouter, openai, anthropic, gemini, litellm, codex, opencode_go, opencode_zen, claude_subscription. See Provider types.
providers.<name>.base_url	string	per type	API endpoint. Required for openai_compat, ollama, lmstudio and litellm.
providers.<name>.api_key	string	none	API key. Prefer api_key_env or ${VAR} to keep secrets out of files.
providers.<name>.api_key_env	string	none	Name of an environment variable holding the API key, read at startup. api_key wins when both are set.
providers.<name>.headers.<header>	string	none	Extra HTTP header sent with every request to this provider.
providers.<name>.timeout	duration	unset	Timeout for each request attempt, covering headers, body and streaming. Unset means no overall request timeout, and a 30s response-header timeout applies instead. Setting a timeout bounds the whole request, including headers, and replaces the header timeout. Retries replay the request, so one call can take up to retry.max_attempts times timeout plus backoff.
providers.<name>.compaction_fallback_after	duration	5m anthropic, 60m claude_subscription, 30m openai and codex, 0 otherwise	Idle time after which compaction may use models.profiles.<profile>.compaction_fallback. 0 disables; negative is invalid. See Context management.
providers.<name>.codex.min_request_interval	duration	0s	codex only. Minimum gap between consecutive requests; 0 disables pacing.
providers.<name>.codex.transport	string	http	codex only. http, or websocket (experimental; failures return an error with no HTTP fallback).

Provider types

Type	base_url	Credentials	Notes
openai_compat	required	api_key or api_key_env optional	Any OpenAI chat-completions compatible server.
ollama	required, e.g. http://localhost:11434/v1	optional	Ollama through its OpenAI-compatible API.
lmstudio	required, typically http://127.0.0.1:1234/v1	optional	LM Studio's local server.
litellm	required	optional	LiteLLM gateway. A 429 without Retry-After has its delay parsed from the body; budget-exhaustion 429s are not retried.
openrouter	default https://openrouter.ai/api/v1	required	OpenRouter gateway.
openai	default https://api.openai.com/v1	required	Native OpenAI API.
anthropic	default https://api.anthropic.com/v1	required	Native Anthropic API. Steiner appends /messages to base_url, so a custom base_url must include the /v1 suffix.
opencode_go	default https://opencode.ai/zen/go/v1	required	OpenCode Go gateway. Create the key in your OpenCode account.
opencode_zen	default https://opencode.ai/zen/v1	required	OpenCode Zen gateway.
codex	optional	ignored	OpenAI Codex subscription. Sign in with steiner login codex; authentication uses that OAuth login, not api_key.
claude_subscription	rejected	rejected	Runs your signed-in claude CLI. base_url, api_key, api_key_env, headers and timeout are rejected. See Claude subscription.
gemini	none	required	Accepted by validation but not supported at runtime; do not use.

"Required" credentials mean one of api_key or api_key_env must be set, or validation fails.

models block

Model definitions, discovery and named profiles. A profile assigns models to roles; a definition (alias) persists settings for one model. Every selection accepts a raw provider/model-id reference, so an alias is only needed to persist settings or give a model a short name. Selection order, aliases and profile usage are in Provider and model setup.

No profile is required. Every declared profile must set default_model, and profiles never inherit from each other. Without --profile, the profile named default is used; if there is none, role assignments are unselected and --model or STEINER_MODEL still sets the orchestrator model.

Path	Type	Default	Description
models.discovery_enabled	bool	true	Discover models from configured providers. false skips provider enumeration and network refresh; the chooser lists configured entries only.
models.definitions.<alias>.provider	string	none	Required. Name of a key in providers.
models.definitions.<alias>.id	string	none	Required. Model ID as the provider's API expects it.
models.definitions.<alias>.params	map[string]any	none	Request parameters sent in the request body, such as temperature, top_p, top_k.
models.definitions.<alias>.extra_params	map[string]any	none	Additional provider-specific body parameters.
models.definitions.<alias>.prompt_suffix	string	none	Text appended to the last user message of each request.
models.definitions.<alias>.vision	bool or null	null	false strips image attachments from requests to this model with a warning. Unset assumes vision support.
models.definitions.<alias>.retry.enabled	bool	true	Retry transient and rate-limit errors. Usage and quota limit errors are never retried.
models.definitions.<alias>.retry.max_attempts	int	5	Total attempts including the first; at least 1.
models.definitions.<alias>.retry.initial_backoff	duration	250ms	Wait before the first retry; greater than zero.
models.definitions.<alias>.retry.max_backoff	duration	5s	Cap on exponential backoff; at least initial_backoff.
models.definitions.<alias>.retry.retry_after_max	duration	60s	Longest Retry-After wait honoured; at least initial_backoff.
models.definitions.<alias>.prompts.system	string	none	Replaces the default system prompt for this model.
models.definitions.<alias>.prompts.system_suffix	string	none	Appended after the default preamble.
models.definitions.<alias>.prompts.compaction	string	none	Replaces the compaction prompt for this model.
models.definitions.<alias>.advanced.limits.context_window	int	resolved from metadata	Context window in tokens. An explicit value wins over discovered metadata. If nothing resolves, 32768 is used.
models.definitions.<alias>.advanced.limits.max_output_tokens	int	resolved from metadata	Maximum tokens per response. If nothing resolves, 4096 is used.
models.definitions.<alias>.advanced.limits.compaction_threshold_tokens	int	70% of the context window	Prompt-token count that triggers compaction; greater than zero and below the resolved context window.
models.definitions.<alias>.advanced.reasoning_echo_back	bool or null	resolved from metadata	Whether the model's reasoning content is echoed back in later requests.
models.definitions.<alias>.advanced.transport	string	auto	auto, openai_compat or anthropic. auto follows models.dev metadata when available, otherwise the provider type. An explicit value overrides metadata.
models.definitions.<alias>.advanced.reasoning.effort	string	unset	Reasoning effort to request, as the provider names it (for example low, high). Unset sends no reasoning field. Must not be blank and must appear in supported_efforts when that is set.
models.definitions.<alias>.advanced.reasoning.supported_efforts	[]string	discovered	Efforts this model accepts, used to validate effort and fill the /model picker. No blank or duplicate values. Unset uses models.dev data.
models.definitions.<alias>.advanced.codex.use_max_context_window	bool	false	codex providers only (rejected otherwise). Use the catalogue's max_context_window when positive, else its context_window.
models.profiles.<profile>.default_model	string	none	Required in every declared profile. A model reference, also the fallback for roles without their own assignment.
models.profiles.<profile>.advisor	string	unset	Model for advisor calls; falls back to default_model.
models.profiles.<profile>.compaction_fallback	string	unset	Model for compaction when the prompt cache is presumed cold. Needs a positive compaction_fallback_after on the main model's provider. Unset disables it; it does not fall back to default_model.
models.profiles.<profile>.sub_agents.<type>	string	unset	Model per agent type: explore, research, code, documentor, evaluate, sanity_check, review, vision. Empty or unset assignments use default_model, except vision, which requires an explicit assignment.
models.profiles.<profile>.oneshot.<phase>	string	unset	Model per oneshot phase: plan, implement, review. Unset phases use default_model.
models.profiles.<profile>.workflow_handoff.<destination>	string	unset	Model per handoff destination: implement, review, build. Unset destinations use default_model.

The vision sub-agent type needs a vision-capable model; the vision tool is not registered unless sub_agents.vision is set. See Sub-agent delegation.

limits block

Runtime limits for turns, tokens, timeouts and tool concurrency. The caps apply to the parent run; tool_output_max_bytes, max_parallel_tools and the tool timeouts also apply to each sub-agent. See Tools for tool behaviour.

Path	Type	Default	Description
limits.max_turns	int	50	Maximum agent-loop turns before the run stops. 0 removes the limit; negative is invalid.
limits.max_tokens	int	500000	Maximum output (completion) tokens generated across a run before it stops; at least 1. Input tokens are not counted.
limits.model_call_timeout	duration	10m	Longest a single provider call may take. 0 disables it. Tool work and sub-agent calls inside a turn are not covered.
limits.tool_timeout_default	duration	30s	Timeout for any tool without an entry in tool_timeouts; greater than zero. Also bounds MCP calls.
limits.tool_timeouts.<tool>	duration	see below	Per-tool timeout; greater than zero. Defaults: bash 600s, read 5s, grep 30s, ls 5s. For an MCP tool the key is its registered name mcp__<server>__<tool>, as shown by steiner tools. The bash entry caps the built-in Bash tool: whole seconds, at most 600s.
limits.tool_output_max_bytes	int	65536	Output kept from one tool call; the rest is truncated. At least 1.
limits.max_parallel_tools	int	4	Parallel-safe tool calls run concurrently in one turn, including built-in read-only tools, LSP tools and MCP tools with a read-only hint. At least 1; 1 is serial. Separate from sub_agent.max_parallel.

sandbox block

Bubblewrap sandbox for bash and subprocess tools. What the sandbox isolates, how approvals interact with it and platform limits are in Sandboxing.

Path	Type	Default	Description
sandbox.enabled	bool	true	Enable the sandbox. --unsafe forces false.
sandbox.warning_on_unsupported_platform	bool	true	Warn when the sandbox is unavailable or bypassed.
sandbox.env_passthrough	[]string	[]	Extra host environment variable names passed to sandboxed processes. A trailing * matches a prefix (MYAPP_*). Entries must be non-empty, contain no =, and * may only be the last character. A bare * is rejected; use env_passthrough_all.
sandbox.env_passthrough_all	bool	false	Pass the full host environment, credentials included, with no filtering. Always prints a warning while the sandbox is enabled, regardless of warning_on_unsupported_platform.
sandbox.bind_host_cache	bool	false	Bind the real ~/.cache read-write instead of a sandbox-private cache directory.
sandbox.host_mounts.<index>.path	string	none	Host path to mount; ~/ expands.
sandbox.host_mounts.<index>.mode	string	read-only	rw for writable access; anything else mounts read-only.

permissions block

Opt-in host capabilities for the sandbox. See Sandboxing.

Path	Type	Default	Description
permissions.docker	bool	false	true leaves the Docker socket reachable inside the sandbox. false masks it and unsets DOCKER_HOST.

sub_agent block

Delegated child agents. Behaviour, agent types and worktrees are in Sub-agent delegation. Per-type models are set in models.profiles.<profile>.sub_agents.

Path	Type	Default	Description
sub_agent.enabled	bool	true	false removes all delegation tools. The constraints below apply only when enabled.
sub_agent.orchestration_level	string	standard	standard or low. low omits the delegation-steering sections of the system prompt.
sub_agent.max_turns	int	120	Turn cap for each child run; at least 15.
sub_agent.max_tokens	int	400000	Maximum output (completion) tokens per child run; at least 1.
sub_agent.max_parallel	int	3	Session-wide cap on running children. Extra spawns queue; more than twice this many outstanding (running plus queued) fail. At least 1.
sub_agent.max_follow_ups	int	5	Follow-up resumes allowed per child; at least 1. Each resume extends the child's turn budget.

advisor block

Optional advisor tool: a stronger model reviews the conversation and returns guidance. The model is chosen by models.profiles.<profile>.advisor. See Advisor.

Path	Type	Default	Description
advisor.enabled	bool	false	Enable the advisor tool.
advisor.max_uses_per_run	int	3	Advisor calls allowed per session; at least 1 when enabled.
advisor.max_uses_per_sub_agent	int	1	Advisor calls allowed per child agent; at least 1 when enabled.
advisor.max_tokens	int	unset	Output-token ceiling for advisor calls; greater than zero when set.
advisor.timeout	duration	180s	Timeout for each advisor request attempt. Overrides providers.<name>.timeout for advisor calls only; greater than zero when set.

oneshot block

Closeout settings for autonomous steiner oneshot runs. Per-phase models are set in models.profiles.<profile>.oneshot. See Oneshot.

Path	Type	Default	Description
oneshot.auto_pr	bool	false	Let closeout push the branch and open a PR or MR after a passing review.

desktop_notifications block

Desktop notifications from the interactive TUI. They fire when a tool approval is required and when a workflow handoff is requested; there is no run-completion notification. See Desktop notifications.

Path	Type	Default	Description
desktop_notifications.enabled	bool	false	Enable notifications.
desktop_notifications.duration	int	0	Display time in seconds. 0 leaves the notification until dismissed; negative is invalid.

update_check block

Background check for a newer Steiner release at startup. It runs in the interactive TUI only. See Installation.

Path	Type	Default	Description
update_check.enabled	bool	true	Enable the startup check.
update_check.interval_hours	int	6	Minimum hours between checks; 0 checks on every startup; negative is invalid.

tools block

Custom subprocess tools registered next to the built-ins. The map key is the tool name the model sees. Names reserved by built-ins (read, glob, grep, ls, bash, display_file, mutate, fetch_url, workflow_handoff) are rejected. The stdin/stdout contract a tool must follow is in Tools.

Path	Type	Default	Description
tools.<name>.exec	string	none	Required. Executable to run; ~/ expands.
tools.<name>.subcommand	string	none	Passed to exec as its first argument.
tools.<name>.description	string	none	Description shown to the model.
tools.<name>.parameters	map[string]any	none	JSON Schema fragment for the tool's input.
tools.<name>.timeout	duration	none	Required; greater than zero.

project_context block

Extra files added to the system prompt, under a byte budget. See Context management.

Path	Type	Default	Description
project_context.max_bytes	int	8000	Byte budget for extra context files; files are truncated or skipped to fit. At least 1.
project_context.max_tokens	int	none	Deprecated alias of max_bytes (multiplied by 4); max_bytes wins when both are set.
project_context.extra_files	[]string	none	Files to include, relative to the project root. A path resolving outside the root is an error; missing files and directories are skipped.
project_context.ignore_files	[]string	none	Entries of extra_files to skip. An ignore entry matches a file by exact path, by base name, or as a trailing path suffix. Globs and directories do not match.

paths block

Filesystem access limits for file tools, including each sub-agent's tools. Exclusion semantics are in Tools.

Path	Type	Default	Description
paths.project_root_only	bool	true	Restrict tools to paths inside the project root.
paths.writable_paths	[]string	[]	When set, an allowlist that restricts mutation tools to these paths, including inside the project root. Outside the root it only adds to what project_root_only permits once that restriction is lifted by approval. May be inside or outside the project root; ~/ expands.
paths.blocked_paths	[]string	[]	Paths always denied, even inside the project root; ~/ expands.
paths.exclude_paths	[]string	none	Paths excluded from searches and directory walks. Relative entries are anchored to the workspace root; ~/ expands.
paths.exclude_patterns	[]string	none	Globs excluded from searches and directory walks.

logging block

Session log files. Treat log files as sensitive: they can contain prompts and tool output. The --log-file flag overrides logging.file and enables logging for the run.

Path	Type	Default	Description
logging.enabled	bool	false	Write the session log to logging.file.
logging.level	string	info	trace, debug, info, warn or error.
logging.file	string	~/.local/share/steiner/steiner.log	Log path; required, ~/ expands. JSONL, appended across runs and rotated by size.
logging.thinking_chunk	bool	false	Include reasoning chunks in the log.
logging.assistant_chunk	bool	false	Include streamed assistant chunks, which duplicate the completed assistant message record.
logging.compaction_log_file	string	none	Separate log of compaction events; ~/ expands.

diagnostics block

Structured JSONL capture for measuring cache hit rates, provider calls, tool runs and TUI cost. Disabled by default and independent of logging: the directory is never derived from logging.file. Each record carries run_id, build_sha and dirty. Streams that write to disk are listed below; see Usage for cache statistics.

Path	Type	Default	Description
diagnostics.enabled	bool	false	Master switch. When false nothing is created.
diagnostics.dir	string	$XDG_STATE_HOME/steiner/diagnostics, else ~/.local/state/steiner/diagnostics	Directory with one file per stream. Required when enabled; must not resolve inside the project root; ~/ expands.
diagnostics.retention_days	int	30	Records older than this are dropped when the writer opens; greater than zero when enabled.
diagnostics.streams.cache	bool	false	Prompt-cache observations (cache.jsonl).
diagnostics.streams.provider	bool	false	One record per model call, whatever the outcome (provider.jsonl).
diagnostics.streams.tool	bool	false	Tool execution and delegation traces (tool.jsonl).
diagnostics.streams.tui	bool	false	Interactive TUI timings, counts and sizes, never content (tui.jsonl).
diagnostics.streams.internal	bool	false	TUI developer error entries (internal.jsonl); sanitised and capped, but may still carry prompt or tool content.
diagnostics.capture_bodies	bool	false	Capture full message, tool and block content instead of bounded fields. Expensive, and captures prompts.

context_management block

Baseline context management. Applies to the parent and to each sub-agent. See Context management.

Path	Type	Default	Description
context_management.read_annotations	bool	true	Annotate a read when an identical full read (same file_hash and range) is already in the carried conversation. false keeps full read results.

search block

Backend for the web_search tool. Setup is in Web search. The key fields can be filled from the unprefixed environment variables in Environment variables when empty.

Path	Type	Default	Description
search.backend	string	unset	searxng, kagi or brave. Selects which fields below are required.
search.searxng_url	string	none	SearXNG base URL. Required for searxng.
search.kagi_api_key	string	none	Kagi API key. Required for kagi.
search.brave_api_key	string	none	Brave Search API key. Required for brave.

mcp block

Model Context Protocol servers. MCP is on by default, but each server must be enabled individually. Approval behaviour and TUI surfaces are in MCP.

Path	Type	Default	Description
mcp.enabled	bool	true	Master switch for the MCP client.
mcp.servers.<server>.enabled	bool	false	Start this server. The name must not contain __.
mcp.servers.<server>.transport	string	stdio	stdio (subprocess) or http.
mcp.servers.<server>.command	string	none	Executable. Required for stdio; must be empty for http.
mcp.servers.<server>.args	[]string	none	Command arguments; must be empty for http.
mcp.servers.<server>.env.<key>	string	none	Extra environment variable for the server process; must be empty for http.
mcp.servers.<server>.url	string	none	http or https endpoint. Required for http; must be empty for stdio.
mcp.servers.<server>.headers.<header>	string	none	Header sent with every request; must be empty for stdio. Names reserved by the MCP SDK (Content-Type, Accept, Mcp-Protocol-Version, Mcp-Session-Id, Last-Event-Id, Mcp-Method, Mcp-Name, and the Mcp-Param- prefix) are rejected. Use ${VAR} for bearer tokens.
mcp.servers.<server>.approval	string	ask	ask, allow or deny. deny registers no tools.
mcp.servers.<server>.trust_annotations	bool	false	Honour a tool's readOnlyHint annotation to skip approval.
mcp.servers.<server>.connect_timeout	duration	15s	Time allowed to connect. Unset or 0 uses 15s; negative is invalid.
mcp.servers.<server>.allowed_tools	[]string	none	Allowlist of the server's native tool names. Unset means no restriction; [] registers nothing.
mcp.servers.<server>.blocked_tools	[]string	none	Native tool names to drop, applied after allowed_tools.
mcp.servers.<server>.sub_agents	[]string	none	Agent types allowed to use this server's tools. Unset or [] grants none. Valid types: explore, research, code, documentor, evaluate, sanity_check, review, vision.

lsp block

Optional language servers for code intelligence tools. Disabled by default; servers are never installed by Steiner. Validation runs only when lsp.enabled is true. Server setup is in LSP.

Path	Type	Default	Description
lsp.enabled	bool	false	Enable LSP tools.
lsp.idle_timeout	duration	5m	Stop an idle server after this long; greater than zero.
lsp.request_timeout	duration	10s	Timeout for one request; greater than zero.
lsp.ready_timeout	duration	30s	Wait for a server to finish indexing before answering anyway; greater than zero.
lsp.ready_grace_period	duration	2s	Grace period for servers that never report progress; greater than zero.
lsp.diagnostics_window	duration	2s	Time to collect published diagnostics after opening a file; greater than zero.
lsp.max_results	int	200	Maximum results per navigation query; greater than zero.
lsp.cache_dir	string	user cache directory	Base directory for server state; ~/ expands.
lsp.servers.<name>.enabled	bool	false	Enable this server. The fields below are validated only for enabled servers.
lsp.servers.<name>.command	string	none	Required. Server executable.
lsp.servers.<name>.args	[]string	none	Command arguments.
lsp.servers.<name>.env.<key>	string	none	Extra environment variable for the server process.
lsp.servers.<name>.file_extensions	[]string	none	Required, non-empty, each starting with .. Two enabled servers cannot claim the same extension (case-insensitive).
lsp.servers.<name>.root_markers	[]string	none	Required, non-empty. Files or directories that identify a workspace root.
lsp.servers.<name>.initialization_options	map[string]any	none	Options passed to the server in the initialize handshake.

modes block

Execution mode at startup. Mode behaviour and switching are in Execution modes.

Path	Type	Default	Description
modes.default	string	build	plan or build. plan limits project edits to .steiner/plans and .steiner/security.

tui block

Interactive terminal UI settings; ignored outside the TUI. See Interactive mode.

Path	Type	Default	Description
tui.fps	int	60	Renderer frame rate, 1 to 120. Higher values reduce input latency at the cost of CPU.
tui.quota_refresh_interval	duration	5m	Age after which a finished run refetches subscription usage for accounts it used. 0 disables only this refresh; negative is invalid. See Usage.

Examples

Minimal local setup with Ollama:

providers:
  local:
    type: ollama
    base_url: http://localhost:11434/v1
models:
  profiles:
    default:
      default_model: local/qwen2.5-coder:14b

Cloud provider (Anthropic):

providers:
  cloud:
    type: anthropic
    base_url: https://api.anthropic.com/v1
    api_key_env: ANTHROPIC_API_KEY
models:
  profiles:
    default:
      default_model: cloud/claude-sonnet-4-5