Configuration reference

Generated from the pydantic settings model — do not edit by hand.

Keys live in ~/.daari/config.yaml (nested YAML), can be overridden per-project in .daari.yaml, and every key is also settable via environment variable: DAARI_<SECTION>__<KEY> (double underscore per nesting level).

Key	Type	Default	Description
server.host	str	'127.0.0.1'
server.port	int	11435
server.api_key	str	list[str]	''
server.dangerously_permit_weak_or_unset_api_key	bool	False	When true, daari serve allows an empty or denylisted master key (sk-1234 / changeme / daari-local / …) for hermetic tests and local sandboxes (#1320). Default false refuses before bind. Env: DAARI_SERVER__DANGEROUSLY_PERMIT_WEAK_OR_UNSET_API_KEY.
server.virtual_keys.enabled	bool	True
server.virtual_keys.path	str	'~/.daari/auth/virtual-keys.sqlite3'
server.virtual_keys.backend	Literal	'sqlite'	sqlite (default) or postgres (observability.postgres_url) so keys and teams resolve across replicas (#544). Env: DAARI_SERVER__VIRTUAL_KEYS__BACKEND.
server.max_body_bytes	int	10485760	Hard cap on inbound request body size (#933). Oversized requests return 413 before the body is buffered. 0 disables the cap. File/audio upload routes may use a higher floor so files.max_total_bytes still applies. Env: DAARI_SERVER__MAX_BODY_BYTES.
server.header_policy.enabled	bool	False	When true, evaluate required/deny/allow before require_api_key. Env: DAARI_SERVER__HEADER_POLICY__ENABLED.
server.header_policy.required	list	[]	Header names that must be present and non-empty.
server.header_policy.deny	list	[]	Denylist rules (exact and/or regex) evaluated against headers.
server.header_policy.allow	dict	{}	Optional allowlists keyed by header name. When the header is present, its value must be one of the listed strings.
server.tls.cert_file	str	''	PEM certificate path (or secret:// ref). Env: DAARI_SERVER__TLS__CERT_FILE.
server.tls.key_file	str	''	PEM private key path or secret:// ref. Env: DAARI_SERVER__TLS__KEY_FILE.
server.tls.client_ca	str	''	Optional client CA path/ref; when set, require a valid client cert (mTLS). Env: DAARI_SERVER__TLS__CLIENT_CA.
server.cors_origins	list	[]	Browser Origin allowlist for CORS (#938). Empty disables CORS middleware. When set, enables ACAO for listed origins, Authorization header, and OPTIONS preflight. Env: DAARI_SERVER__CORS_ORIGINS (JSON list, e.g. '["http://127.0.0.1:11437"]').
server.security_headers	bool	True	Attach baseline security headers on every response (#938): X-Content-Type-Options: nosniff, X-Frame-Options: DENY, Referrer-Policy: no-referrer. Env: DAARI_SERVER__SECURITY_HEADERS.
server.sse_keepalive_seconds	float	10.0	Idle seconds between streamed chunks before emitting a keepalive frame for the entire stream lifetime (#972) (SSE comment : keepalive on OpenAI/Anthropic/Responses routes, a blank line on the NDJSON Ollama facade). Keeps proxies and SDK read timeouts from dropping slow streams (including mid-generation pauses). 0 disables.
server.stream_idle_timeout_seconds	float	0.0	If upstream produces no chunk for this many seconds mid-stream, end with an in-band error event instead of hanging (#972). 0 disables (default).
server.graceful_timeout_seconds	float	30.0	Seconds uvicorn waits for in-flight requests (including SSE) after SIGTERM/SIGINT before force-closing (#1104). Wire via daari serve --graceful-timeout or DAARI_SERVER__GRACEFUL_TIMEOUT_SECONDS. Size terminationGracePeriodSeconds above this plus preStop sleep.
auth.key_lockout.enabled	bool	True	When false, invalid attempts are not counted per key prefix. Env: DAARI_AUTH__KEY_LOCKOUT__ENABLED.
auth.key_lockout.max_failures	int	10	Invalid-auth attempts for one key prefix within window_seconds before lockout. 0 disables the counter. Matches auth.max_failures. Env: DAARI_AUTH__KEY_LOCKOUT__MAX_FAILURES.
auth.key_lockout.window_seconds	float	60.0	Sliding window for auth.key_lockout.max_failures. Env: DAARI_AUTH__KEY_LOCKOUT__WINDOW_SECONDS.
auth.key_lockout.lockout_seconds	int	60	How long a tripped key prefix stays locked (Retry-After). Env: DAARI_AUTH__KEY_LOCKOUT__LOCKOUT_SECONDS.
auth.throttle_enabled	bool	True	When false, invalid-key attempts are never rate-limited.
auth.max_failures	int	10	Invalid-key failures per client IP within window_seconds before 429. 0 disables the counter. Env: DAARI_AUTH__MAX_FAILURES.
auth.window_seconds	float	60.0	Sliding window for auth.max_failures. Env: DAARI_AUTH__WINDOW_SECONDS.
auth.exempt_loopback	bool	True	Skip throttling for 127.0.0.1 / ::1 / localhost.
secrets.refresh_ttl_seconds	float	300.0	TTL for re-resolving env-file/exec/keychain secret:// refs at use time. 0 = resolve once at boot (legacy). Default 300s. Env-file refs also re-resolve when the source file mtime changes. Env: DAARI_SECRETS__REFRESH_TTL_SECONDS.
rate_limit.rpm	int	0	Default requests per minute per key (0=unlimited).
rate_limit.tpm	int	0	Default tokens per minute per key (0=unlimited).
rate_limit.model_rpm	int	0	Per-key-per-model RPM. 0 falls back to rpm.
rate_limit.model_tpm	int	0	Per-key-per-model TPM. 0 falls back to tpm.
rate_limit.models	dict	{}	Named per-model-id rpm/tpm/max_in_flight overrides (#1592). Exact id match wins; otherwise longest prefix match.
rate_limit.max_in_flight	int	0	Global in-flight request cap. 0 disables the concurrency gate.
rate_limit.queue_size	int	32	Waiters allowed when in-flight is full; overflow is 503 + Retry-After.
rate_limit.retry_after_seconds	int	1	Retry-After value on 429/503.
rate_limit.fail_open	bool	False	When Redis counters are unreachable, allow requests without counting instead of degrading to the per-replica SQLite backend. Default false (prefer SQLite fallback so limits still apply locally).
models.l3	str	'llama3.2:3b'
models.l4	str	'llama3.1:8b'
models.l5	str	'llama3.1:70b'
models.weights	dict	{}
models.capabilities	dict	{}
models.lifecycle	dict	{}
models.reject_retired	bool	True	When true (default), chat/messages/responses to a model whose catalog lifecycle is retired return HTTP 410 instead of routing (fail-closed).
models.timeout_s	dict	{}	Optional per-tier request timeout in seconds (keys L3/L4/L5). Unset tiers use upstream.local_timeout_seconds.
ollama.base_url	str	'http://127.0.0.1:11434'
systemone.enabled	bool	True	When true, POST /v1/systemone proxies to ollama.base_url and local decisions ids are advertised on GET /v1/models (OpenAI list tag decision and Anthropic capabilities.decision.supported) and Ollama /api/tags with capability decision. When false, the route returns 501 not_implemented and discovery omits those ids.
decisions.cache_enabled	bool	True	When true (default), identical typed decision requests are served from the L0 exact cache. Requires cache.l0.enabled. Clients can bypass with X-Daari-No-Cache: true.
decisions.local_models	list	[]	Extra local decisions model ids that extend the built-in catalog (clef, clef-flash, nimble, tev1, drex-1.5, microsoft-decision-1). Operators can add future Ollama judge models without a daari release. Advertised on OpenAI + Anthropic-shaped /v1/models and /api/tags when systemone.enabled.
mlx.enabled	bool	False
mlx.base_url	str	'http://127.0.0.1:11440'
mlx.models	dict	{}
asr.base_url	str	''	OpenAI-compatible ASR base URL, including /v1 (vLLM, whisper.cpp server, or another local pool member). Empty leaves POST /v1/audio/transcriptions unconfigured.
asr.model	str	''	Optional model name sent to the ASR server. When set, it replaces the client model so a local server always sees its own id.
asr.frontier_fallback	bool	False	When true and asr.base_url is empty, forward one transcription to the configured frontier base if frontier.enabled and a key is present. Default false so audio is never uploaded to a cloud endpoint implicitly.
ocr.base_url	str	''	OpenAI-/Mistral-compatible OCR API root, including /v1. When set, POST /v1/ocr forwards to {base_url}/ocr. Empty falls through to ocr.vision_model or frontier L6.
ocr.model	str	''	Optional default OCR model id when the client omits model. Used for local base_url and frontier passthrough.
ocr.vision_model	str	''	When base_url is empty, run OCR via local Ollama multimodal chat at ollama.base_url using this vision model (image_url / image data URIs). Empty skips the local vision path.
tts.base_url	str	''	OpenAI-compatible TTS base URL, including /v1 (openedai-speech, Kokoro-FastAPI, or similar). Empty leaves POST /v1/audio/speech unconfigured.
tts.model	str	''	Optional model name sent to the TTS server. When set, it replaces the client model so a local server always sees its own id.
tts.voice	str	''	Optional default voice when the request omits voice (OpenAI alloy/echo/… or the local server's voice id).
cache.l0.enabled	bool	True
cache.l0.path	str	'~/.daari/cache/l0'
cache.l0.ttl_seconds	float	0.0
cache.l1.enabled	bool	True
cache.l1.path	str	'~/.daari/cache/l1'
cache.l1.similarity_threshold	float	0.88
cache.l1.draft_threshold	float	0.75
cache.l1.max_entries	int	1000
cache.l1.embedding_model	str	'nomic-embed-text'
cache.l1.ttl_seconds	float	0.0
cache.l1.embed_cache_size	int	512
cache.l1.normalize_inputs	bool	True
cache.l1.verify	Literal	'lexical'	Second-stage check before serving a semantic hit, because a cosine threshold alone cannot separate a paraphrase from a near-miss. none serves any hit above the threshold; lexical (default) vetoes hits whose numbers, units, or negation differ; model additionally asks a local model to confirm equivalence.
cache.l1.shadow_sample_rate	float	0.05
cache.backend	Literal	'disk'
cache.redis_url	str	'redis://127.0.0.1:6379/0'	Redis URL for L0/L1 and shared counters when cache.backend=redis. Localhost may use redis://; remote hosts should use rediss:// (TLS). Optional CA pass-through: rediss://user:pass@redis.example:6379/0?ssl_ca_certs=/etc/ssl/certs/ca.pem. daari doctor warns on cleartext remote URLs (redis_tls, #1679).
cache.redis_prefix	str	'daari:l0:'
cache.redis_l1_prefix	str	'daari:l1:'
cache.redis_timeout_seconds	float	2.0	socket_connect_timeout and socket_timeout for every Redis client (rate-limit counters, L0/L1 cache, budget-alert dedupe). Keeps a hung Redis from blocking gateway requests indefinitely (#463).
routing.prefer	Literal	'balanced'
routing.confidence_threshold	float	0.7
routing.category_policies	dict	{}
routing.max_tier_for_chat	Optional	None
routing.latency_budget_ms	int	0
routing.warm_model_preference	bool	True
routing.ttft_aware	bool	False	When true, after the heuristic pick, prefer a faster local tier (L3..heuristic) with enough recent TTFT samples and a lower configured percentile. Default off.
routing.ttft_percentile	float	0.95	TTFT histogram percentile used when ttft_aware is true.
routing.ttft_min_samples	int	20	Minimum TTFT samples per tier before ttft_aware may prefer it.
routing.learned_router	bool	False
routing.reasoning_effort_escalation	bool	False
routing.compact_to_fit.enabled	bool	False	When true, chat history is compacted to max_messages / max_tokens before L6. System messages and tool-call payloads are never deleted; if the remainder still exceeds the cap, history is left oversized (fail closed). A successful trim records tokens_before / tokens_after on the compact_to_fit trace step and increments compact_to_fit_tokens_dropped on /v1/daari/stats. Default off. Editable via GET/PATCH /v1/daari/config ownership (config editor).
routing.compact_to_fit.max_messages	int	32	Keep at most this many messages when compact_to_fit is on. Positive integer. Editable via config editor ownership.
routing.compact_to_fit.max_tokens	int	0	Approximate token budget (chars/4) when compact_to_fit is on. 0 disables the token cap. Non-negative integer. Editable via config editor ownership.
routing.stall_escalation.enabled	bool	False	When true, N identical tool calls in the last window, or N consecutive error tool results, escalate the chosen tier by one. Default off.
routing.stall_escalation.repeats	int	3	Identical calls or consecutive error results required to stall.
routing.stall_escalation.window	int	6	How many recent tool calls are inspected for identical repeats.
routing.phase_routing.enabled	bool	False	When true, the last window of tool-call names classifies the turn as explore / implement / verify and adjusts the heuristic tier. Default off.
routing.phase_routing.window	int	6	How many recent tool-call names are classified for phase.
routing.phase_routing.explore	int	str	-1
routing.phase_routing.implement	int	str	0
routing.phase_routing.verify	int	str	0
routing.decision_classifier.enabled	bool	False	When true, Ask (and optionally Agent) turns call Ollama /v1/systemone before the heuristic tier pick. Default off. Editable via GET/PATCH /v1/daari/config ownership (config editor).
routing.decision_classifier.model	str	'nimble'	Decision model name proxied to Ollama /v1/systemone. Editable via config editor ownership.
routing.decision_classifier.timeout_seconds	float	5.0	Classifier request timeout; on timeout fall back to heuristics. Editable via config editor ownership.
routing.decision_classifier.agent_turns	bool	False	When true, also classify agent/tool turns. Default false (Ask only). Editable via config editor ownership.
routing.session_affinity	bool	False	When true, a tool-result continuation or an unchanged user-turn prefix reuses the session's prior tier instead of re-running rules. A new human turn re-routes. Default off.
routing.session_affinity_ttl_seconds	float	1800.0	How long a session pin (and the session cost-avoided rollup) is reused. 0 keeps the pin until process restart (in-process) or without a Redis TTL when cache.backend=redis. Ignored unless session_affinity is true.
routing.classify_user_turn	bool	False	When true, a tool-result continuation (no new user text) reuses the previous profile's category/complexity. Phase routing and stall escalation still inspect tool history. A new user message re-profiles. Default off.
routing.classify_user_turn_agents	bool	True	When classify_user_turn is false, still reuse profiles on tool-result continuations for agent User-Agents (cursor, claude-code, claude code, codex). Set false to disable the shortcut. Ignored when classify_user_turn is true (applies to every client). Default on.
routing.harness_aware_profile	bool	True	When true, complexity/category ignore system catalogs and recognized Codex/Claude Code harness blocks (system-reminder, environment_context, recommended_plugins, user_instructions, environments_instructions, agents.md envelope). prompt_tokens_est still counts the full request (#418, #541). Default on (no-op without those markers).
routing.context_window_escalation	bool	True	When true, pick a higher local tier before the first hop if the prompt estimate exceeds that tier's known context window (#385).
routing.context_window_escalation_buffer	float	0.95	Escalate when estimated tokens exceed window * buffer.
routing.context_windows	dict	{'L3': 8192, 'L4': 32768, 'L5': 131072}	Known context windows (tokens) per local tier. Missing = unknown, left alone.
routing.shadow_sample_rate	float	0.0	Fraction of local-tier responses replayed in the background at shadow_compare_tier to measure tier divergence. 0 disables.
routing.shadow_compare_tier	Literal	''	Tier to replay sampled requests at. Empty = highest configured local tier; L6 requires shadow_daily_usd > 0.
routing.shadow_daily_usd	float	0.0	Daily spend cap for L6 shadow replays. 0 forbids L6 shadow runs.
routing.org_pool.enabled	bool	False
routing.org_pool.base_url	str	''
routing.org_pool.model	str	''
routing.org_pool.tier	str	'L5-org'
routing.local_pool.strategy	str	'least_outstanding'	Host pick: least_outstanding or round_robin. Warm models still win ties.
routing.local_pool.health_interval_seconds	float	15.0	Background health-check interval. Requests use the last snapshot.
routing.local_pool.frontier_fallback	bool	False	When true and every local backend for the chosen tier is down or circuit-open, escalate to L6 instead of raising BackendUnavailable. Respects no_frontier, allowlists, budgets, and PII scrub. Default false so outages stay a hard 503 unless opted in (#846).
routing.local_pool.backends	list	[]
frontier.enabled	bool	False
frontier.provider	str	'openai'
frontier.model	str	'gpt-6-luna'
frontier.confidence_threshold	float	0.7
frontier.base_url	str	'https://api.openai.com/v1'
frontier.providers	list	[]	Ordered L6 failover chain. Optional per-entry timeout_s, retry_attempts, and retry_backoff_s fall back to upstream.frontier_timeout_seconds / upstream.retry when unset.
frontier.param_compat	dict	{}	Per-model frontier parameter compatibility overrides. Merged over the builtin table (e.g. gpt-6-astra); empty keeps defaults.
frontier.daily_budget_usd	float	0.0
frontier.monthly_budget_usd	float	0.0
frontier.soft_budget_ratio	float	0.8
frontier.scrub_pii	bool	False
frontier.price_per_1k_tokens	float	0.002
frontier.slim_prompts	bool	True
frontier.max_history_messages	int	8
frontier.prompt_cache	bool	True
frontier.compress_context	bool	False
frontier.compress_target_ratio	float	0.6
tools.unknown	str	'deny'
tools.allow	list	['git status', 'git diff', 'pytest', 'eslint *']
tools.block	list	['rm *', 'curl *\| sh', '*> /dev/*']
tools.timeout_seconds	float	30.0
context.enabled	bool	True
context.path	str	'~/.daari/context/commands'
usage.enabled	bool	True
usage.path	str	'~/.daari/usage/ledger.sqlite3'
usage.frontier_price_per_1k_tokens	float	0.002	Flat fallback rate used to estimate what locally-served tokens would have cost on a frontier model. Applies only to models absent from pricing.models, and ignores input/output direction.
usage.spend.enabled	bool	False	Write one spend row per completed request (timestamp, key, team, tokens, cost, cost avoided). Off keeps the day ledger's write volume.
usage.spend.path	str	'~/.daari/usage/spend.sqlite3'	SQLite path for per-request spend rows. Ignored when observability.backend is postgres.
files.enabled	bool	True
files.path	str	'~/.daari/files'
files.backend	Literal	'sqlite'	sqlite (disk index + .bin files, default) or postgres (metadata + BYTEA content via observability.postgres_url) for multi-replica fleets (#465).
files.max_bytes	int	104857600	Maximum upload size in bytes for POST /v1/files.
files.retention_days	int	0	Fallback expiry for files without expires_after (#456). 0 keeps files forever until an explicit expires_after.
files.max_total_bytes	int	0	Hard cap on aggregate stored bytes (#456). 0 disables the cap. Uploads that would exceed it return 413.
batches.enabled	bool	True
batches.path	str	'~/.daari/batches/jobs.sqlite3'
batches.backend	Literal	'sqlite'	sqlite (default) or postgres (observability.postgres_url) so batch jobs are readable/cancellable across replicas (#465).
batches.claim_ttl_seconds	int	90	How long a replica's drain claim stays exclusive before another replica may reclaim a crashed worker (#465). Ignored for sqlite.
batches.yield_to_interactive	bool	True	When true, the batch worker waits while interactive HTTP requests are in flight before dispatching the next item (#444).
batches.idle_poll_seconds	float	0.25	How often to re-check interactive load while yielding.
responses.backend	Literal	'sqlite'	sqlite (default, path next to traces) or postgres (observability.postgres_url) so store:true / previous_response_id / background polling work across replicas (#481).
responses.retention_days	int	0	Delete stored responses older than this many days (#497). 0 keeps them forever.
idempotency.enabled	bool	True	When true, honor Idempotency-Key on chat completions, Responses, modality routes, and decisions. A missing header is always a no-op.
idempotency.backend	Literal	'sqlite'	sqlite (default, path next to traces) or postgres (observability.postgres_url) for multi-replica fleets.
idempotency.ttl_seconds	int	86400	How long completed idempotency records are kept (default 24h).
idempotency.wait_seconds	float	60.0	How long an in-flight duplicate waits for the first request.
pricing.models	dict	{'gpt-4o': {'input_per_1m': 2.5, 'output_per_1m': 10.0, 'cached_input_per_1m': 1.25, 'cache_write_1h_per_1m': None, 'cache_write_per_1m': No…	Per-model, per-direction USD rates per 1M tokens. Keys match on longest prefix, so gpt-4o also prices gpt-4o-2024-08-06 and a vendor prefix (anthropic.claude-fable-5-1) resolves the same way. Models absent here fall back to usage.frontier_price_per_1k_tokens; run daari doctor to list models being billed at the fallback rate.
upstream.local_timeout_seconds	float	120.0	Request timeout for local backends (Ollama, MLX). Generous because a large local model on a cold start can be genuinely slow.
upstream.frontier_timeout_seconds	float	90.0	Request timeout for frontier (L6) providers. Lower than local, since a hosted API that has not answered in 90s is usually not going to.
upstream.request_deadline_seconds	float	None	None
upstream.retry.attempts	int	3	Total attempts per upstream call, counting the first. 1 disables retries. Only transient failures are retried (408, 429, 5xx, connect and read timeouts); a 401 or malformed body fails immediately.
upstream.retry.base_delay_ms	int	200	First backoff, doubled per retry up to max_delay_ms.
upstream.retry.max_delay_ms	int	5000	Ceiling for a single backoff interval.
upstream.retry.jitter	float	0.5	Fraction of each backoff that is randomized, keeping the delay in [d*(1-jitter), d]. Spreads retries from requests that failed together instead of returning them in lockstep.
upstream.pool_max_connections	int	100	Max concurrent connections across the shared upstream httpx pool (Ollama, OpenAI-compat, MLX, frontier, embedder, TTS, ASR).
upstream.pool_keepalive_connections	int	20	Max idle keepalive connections retained in the shared upstream httpx pool. Cuts TCP/TLS handshake cost on repeated local hops.
trace.enabled	bool	True
trace.path	str	'~/.daari/traces/traces.sqlite3'
trace.max_entries	int	200
observability.request_log_max_bytes	int	5242880
observability.request_log_backups	int	3
observability.request_log.backend	Literal	'jsonl'	jsonl (default, pod-local rotated file) or postgres (observability.postgres_url) so fleet replicas share one queryable event table (#1631).
observability.prometheus	bool	True
observability.metrics_port	int	0
observability.otel	bool	False
observability.config_editor	bool	False
observability.backend	Literal	'sqlite'
observability.postgres_url	str	''	Shared Postgres DSN for fleet stores (ledger, virtual keys, batches/files/responses, request log, audit). Localhost/unix sockets may omit TLS; remote DSNs should include sslmode=require (or verify-ca / verify-full). Optional CA: ?sslmode=verify-full&sslrootcert=/etc/ssl/certs/ca.pem. Example: postgresql://daari:daari@db.example:5432/daari?sslmode=require. daari doctor warns on cleartext remote DSNs (postgres_tls, #1679).
observability.postgres_pool_min	int	1
observability.postgres_pool_max	int	4
observability.structured_json_logs	bool	False	Emit gateway request logs as single-line JSON to stdout (containers/SIEM). Default off. Editable via GET/PATCH /v1/daari/config ownership (config editor).
observability.otlp_logs	bool	False
observability.stateless	bool	False
observability.retention.traces_days	int	0
observability.retention.ledger_days	int	0
observability.retention.audit_days	int	0
observability.retention.shadow_days	int	0
observability.retention.tasks_days	int	0
observability.retention.spend_days	int	0	Delete per-request spend rows older than this many days (#709). 0 keeps them forever.
observability.retention.request_log_days	int	0	Delete gateway request-log lines and rotated backups older than this many days (#772). 0 keeps size-only rotation.
observability.retention.batches_days	int	0	Delete batch jobs older than this many days (#1136). 0 keeps them forever.
observability.retention.cache_prune	bool	False	When true, daari prune / daily sweep invoke L0 and L1 prune() using each cache's ttl_seconds (#1136).
learning.enabled	bool	True
learning.path	str	'~/.daari/feedback/feedback.sqlite3'
learning.max_rows	int	20000
learning.auto_tune	bool	False
learning.tuner_min_samples	int	50
learning.capture_examples	bool	False
learning.examples_path	str	'~/.daari/training/examples.sqlite3'
learning.examples_max_rows	int	5000
learning.router_min_samples	int	200
learning.router_model_path	str	'~/.daari/learning/router-model.json'
learning.collective_enabled	bool	False
learning.collective_url	str	''
learning.collective_token	str	''
context_optimizer.enabled	bool	True
context_optimizer.max_history_messages	int	20
context_optimizer.squeeze_whitespace	bool	True
context_optimizer.compact	bool	False
guardrails.enabled	bool	False
guardrails.max_prompt_chars	int	0
guardrails.injection_action	str	'block'
guardrails.block_message	str	'Request blocked by daari guardrail.'
guardrails.input_rules	list	[]
guardrails.output_rules	list	[]
guardrails.stream_mode	Literal	'buffered'	How output guardrails apply to SSE streams. buffered (default) scans the full answer before the first byte; incremental scans with a holdback window and keeps frontier relay eligible.
guardrails.stream_holdback_chars	int	256	Characters held back before emission in incremental stream_mode so a secret spanning two deltas is caught. Ignored when buffered.
guardrails.scan_tool_results	bool	False	When true, scan OpenAI role=tool and Anthropic tool_result message contents with output rules (secrets/PII/deny) before the model hop. System/user/assistant messages are unchanged. Default off.
boundaries.enabled	bool	False
boundaries.mode	Literal	'block'
boundaries.product_name	str	''
boundaries.product_description	str	''
boundaries.allow_topics	list	[]
boundaries.deny_topics	list	[]
boundaries.examples_in	list	[]
boundaries.examples_out	list	[]
boundaries.refuse_message	str	'This assistant can only help with in-product questions.'
boundaries.clear_out_threshold	float	0.85
boundaries.clear_in_threshold	float	0.85
boundaries.local_judge_model	str	None	None
boundaries.quorum_votes	int	2
boundaries.frontier_judge_daily_budget_usd	float	0.5
boundaries.stages_b0	bool	True
boundaries.stages_b1	bool	True
boundaries.stages_b2	bool	True
boundaries.stages_b3	bool	False
boundaries.active_profile	str	''
boundaries.profiles	dict	{}
integrations.sourcegraph.url	str	(required)
integrations.sourcegraph.triggers	list	[]
integrations.ghe.url	str	(required)
integrations.ghe.triggers	list	[]
integrations.gitlab.url	str	(required)
integrations.gitlab.triggers	list	[]
integrations.mcp_servers	list	[]	External MCP servers daari can call. Each entry: id, url, optional token/triggers; optional rpm/tpm (0=unlimited) per-server egress ceilings on tools/call and tools/list (429 + Retry-After + X-RateLimit-Scope mcp_server:; independent of key-family mcp; both apply) (#1620). Per-server rpm/tpm are editable via GET/PATCH /v1/daari/config ownership (config editor; id + rpm/tpm only; secrets omitted) (#1659). Opt-in OBO (#1319): auth_type=oauth2_token_exchange with token_exchange_endpoint, client_id, client_secret; optional audience, scopes, subject_token_type (default access_token).
integrations.mcp_policy.allow	list	[]	MCP tool names (glob) the caller may call. Empty = every tool not denied.
integrations.mcp_policy.deny	list	[]	MCP tool names (glob) the caller may never call. Deny beats allow.
integrations.mcp_policy.servers.allow	list	[]	MCP egress server ids (glob) the caller may reach. Empty = every server not denied.
integrations.mcp_policy.servers.deny	list	[]	MCP egress server ids (glob) the caller may never reach. Deny beats allow.
integrations.mcp_policy.clients.allow	list	[]	MCP client identities (glob) allowed to open the gateway. Empty = every client not denied. Match on clientInfo.name / OAuth client_id / User-Agent.
integrations.mcp_policy.clients.deny	list	[]	MCP client identities (glob) that may never open the gateway. Deny beats allow.
integrations.mcp_policy.require_key_access_defined	bool	False	When true, virtual keys with no MCP grant in metadata.mcp (allow / deny / servers.allow / servers.deny) fail initialize (HTTP 403 JSON-RPC -32003), get an empty tools/list, and tools/call deny on /mcp and /v1/mcp/query (#1352, #1389). Master key and flag-off behavior unchanged. Default false. Editable via GET/PATCH /v1/daari/config ownership (config editor).
integrations.mcp_team_policies	dict	{}	Per-team MCP tool policy keyed by team name; layered on mcp_policy.
integrations.mcp_tasks.enabled	bool	True
integrations.mcp_tasks.long_running_tools	list	['route']
integrations.mcp_tasks.threshold_ms	int	0
integrations.mcp_tasks.path	str	'~/.daari/mcp-tasks'
integrations.mcp_tool_search.enabled	bool	False	When true and the catalog exceeds min_catalog_size, rank tools by embedding similarity and return top_k. Default off — listing is unchanged.
integrations.mcp_tool_search.min_catalog_size	int	40	Catalogs at or under this size are returned unranked.
integrations.mcp_tool_search.top_k	int	40	Maximum tools returned after ranking.
integrations.mcp_list_cache.ttl_ms	int	60000	Client-hint TTL in milliseconds for tools/list _meta.ttlMs. 0 still emits the field; clients may treat it as uncacheable.
integrations.mcp_egress.allow_private_networks	bool	False	When true, MCP egress may POST to loopback, RFC1918, and link-local hosts (docker-compose / lab). Default false rejects those addresses after DNS resolution.
integrations.mcp_egress.failure_threshold	int	3	Consecutive MCP egress failures before the per-server circuit opens (#1203). Matches frontier failure_threshold.
integrations.mcp_egress.cooldown_seconds	float	30.0	Seconds an open MCP egress circuit stays fail-fast before a half-open probe (#1203).
integrations.mcp_egress.server_side_responses	bool	False	When true, Responses may include type: mcp tools that resolve to a configured integrations.mcp_servers id (server_label); daari lists tools, runs egress tools/call on model tool rounds, and continues the turn (#1232). Default false keeps the #1135 honesty 400 for hosted types.
integrations.mcp_egress.server_side_messages	bool	False	When true, Messages may include mcp_servers + tools entries with type: mcp_toolset whose mcp_server_name resolves to a configured integrations.mcp_servers id; daari lists tools, runs egress tools/call on model tool rounds, and continues the turn (#1261). Default false fails closed with an honesty 400.
integrations.mcp_aggregate_egress.enabled	bool	False	When true, /mcp tools/list includes tools from configured integrations.mcp_servers, namespaced as {server_id}__{tool}; tools/call on those names routes through the egress path (SSRF, breaker, guardrails, metering). Denied servers are omitted; per-server list failures degrade without emptying the catalog. Default false keeps the first-party-only catalog (#1294). Editable via GET/PATCH /v1/daari/config ownership (config editor).
integrations.mcp_guardrails.enabled	bool	False
integrations.mcp_guardrails.max_prompt_chars	int	0
integrations.mcp_guardrails.injection_action	str	'block'
integrations.mcp_guardrails.block_message	str	'Request blocked by daari guardrail.'
integrations.mcp_guardrails.input_rules	list	[]
integrations.mcp_guardrails.output_rules	list	[]
integrations.mcp_guardrails.stream_mode	Literal	'buffered'	How output guardrails apply to SSE streams. buffered (default) scans the full answer before the first byte; incremental scans with a holdback window and keeps frontier relay eligible.
integrations.mcp_guardrails.stream_holdback_chars	int	256	Characters held back before emission in incremental stream_mode so a secret spanning two deltas is caught. Ignored when buffered.
integrations.mcp_guardrails.scan_tool_results	bool	False	When true, scan OpenAI role=tool and Anthropic tool_result message contents with output rules (secrets/PII/deny) before the model hop. System/user/assistant messages are unchanged. Default off.
integrations.mcp_oauth.protected_resource	bool	False	When true, serve GET /.well-known/oauth-protected-resource (and the /mcp-scoped variant) and attach a WWW-Authenticate challenge on unauthenticated /mcp requests when API-key auth is on (#1262). Default false leaves the existing 401 body unchanged. Editable via GET/PATCH /v1/daari/config ownership (config editor); secrets stay non-editable.
integrations.mcp_oauth.local_as	bool	False	When true, expose a minimal on-box OAuth authorization server (GET /.well-known/oauth-authorization-server + POST /oauth/token client_credentials + POST /oauth/revoke RFC 7009) that mints short-lived Bearer tokens bound to an existing API/virtual key (#1293, #1475). Enforces scopes_supported on mint and the mcp scope on /mcp; revoked jti values fail verification. Advertises the local issuer in authorization_servers when that list is empty. Default false — discovery-only / external IdP stays the protected_resource path. Editable via config editor ownership.
integrations.mcp_oauth.resource	str	''	Optional public base URL for the resource metadata resource field (defaults to the request base URL). Env: DAARI_INTEGRATIONS__MCP_OAUTH__RESOURCE.
integrations.mcp_oauth.authorization_servers	list	[]	Authorization server issuer URLs advertised in the metadata document.
integrations.mcp_oauth.scopes_supported	list	['mcp']	OAuth scopes advertised for the MCP resource.
integrations.mcp_oauth.token_ttl_seconds	int	900	Access-token lifetime in seconds for local AS mint (#1293). Default 900.
integrations.mcp_oauth.signing_secret	str	''	HS256 secret for minting MCP access tokens (#1293). Empty derives from the primary master key. Env: DAARI_INTEGRATIONS__MCP_OAUTH__SIGNING_SECRET. Not editable via the config editor (secrets stay redacted).
integrations.mcp_openapi_proxy.enabled	bool	False	When true, POST /mcp/proxy lists and calls tools derived from configured OpenAPI specs. Default false.
integrations.mcp_openapi_proxy.specs	list	[]	Allowlisted OpenAPI sources (id + openapi_url required).
integrations.mcp_registry.enabled	bool	False	When true, serve GET /v1/mcp/registry.json listing the built-in /mcp ingress and configured integrations.mcp_servers URLs. Advertisement-only; /mcp still enforces API-key / allowlists. Default false returns 404. Editable via config editor ownership.
integrations.mcp_registry.public_base_url	str	''	Optional public base URL for the built-in daari server entry (defaults to the request base URL). Env: DAARI_INTEGRATIONS__MCP_REGISTRY__PUBLIC_BASE_URL.
enterprise.enabled	bool	False
enterprise.id	str	None	None
enterprise.org_id	str	None	None
enterprise.tenant_id	str	None	None
enterprise.control_plane_url	str	None	None
enterprise.org_token	str	None	None
enterprise.shared_cache_url	str	None	None
enterprise.shared_cache_token	str	None	None
enterprise.shared_cache_require_token	bool	False
enterprise.shared_cache_timeout_seconds	float	1.0
enterprise.shared_cache_max_retries	int	2
enterprise.shared_cache_backoff_seconds	float	0.2
enterprise.shared_cache_path	str	None	None
enterprise.learning_enabled	bool	False
enterprise.learning_url	str	None	None
enterprise.learning_token	str	None	None
enterprise.learning_timeout_seconds	float	0.5
enterprise.learning_sync_seconds	float	300.0
enterprise.learning_path	str	None	None
enterprise.policy_overrides	dict	{}
enterprise.profile	str	'developer'
enterprise.device_id	str	None	None
enterprise.config_signing_secret	str	''
enterprise.policy_sync_url	str	None	None
enterprise.cache.enabled	bool	False
enterprise.cache.share_classes	list	[]
enterprise.cache.no_org_cache_default	bool	False
enterprise.learning.enabled	bool	False
enterprise.learning.upload_prompts	bool	False
enterprise.learning.upload_code	bool	False
enterprise.sso.enabled	bool	False
enterprise.sso.issuer	str	'daari-dev'
enterprise.sso.secret	str	''
enterprise.sso.jwks_url	str	''
enterprise.sso.jwks_urls	list	[]
enterprise.sso.discovery_url	str	''
enterprise.sso.audience	str	''
enterprise.sso.role_claim	str	'role'
enterprise.sso.admin_min_role	str	'admin'
enterprise.sso.mint_virtual_key_on_login	bool	False
enterprise.sso.mapping_claim	str	'groups'
enterprise.sso.key_mappings	dict	{}
enterprise.sso.default_policy	daari.enterprise.config.SsoKeyPolicy	None	None
enterprise.sso.deny_unmapped	bool	False
enterprise.audit_path	str	'~/.daari/audit/audit.sqlite3'
enterprise.audit_backend	Literal	'sqlite'	sqlite (default, per-node file) or postgres (observability.postgres_url) for a single fleet-wide hash chain (#483).
alerts.budget_webhook_url	str	''
alerts.budget_webhook_secret	str	''
alerts.budget_webhook_secret_previous	str	''	Previous budget webhook HMAC secret kept during rotation grace (#1680). Signing always uses budget_webhook_secret; receivers should verify against current or previous. Set via daari alerts rotate-secret budget.
alerts.budget_webhook_secret_rotated_at	str	''	ISO-8601 timestamp when budget_webhook_secret_previous was last set by rotation (#1680). Doctor tips when the previous secret has been in place a long time.
alerts.budget_thresholds	list	[0.8, 1.0]
alerts.member_budget_thresholds	list	[]	Optional ratios of a key's user_daily_usd_cap. When set, a named end user on a team key crossing one of these ratios POSTs a budget webhook with scope=member (#1633). Empty disables member alerts and leaves key/team alerts unchanged. Env: DAARI_ALERTS__MEMBER_BUDGET_THRESHOLDS.
alerts.key_expiry_webhook_url	str	''	Opt-in webhook for virtual keys entering the expiring-soon window (#1546). Empty disables.
alerts.key_expiry_webhook_secret	str	''	HMAC secret for key_expiry webhooks (X-Daari-Signature). Supports secret:// refs (#1546).
alerts.key_expiry_webhook_secret_previous	str	''	Previous key_expiry webhook HMAC secret during rotation grace (#1680). Signing uses key_expiry_webhook_secret; receivers verify against either. Set via daari alerts rotate-secret key-expiry.
alerts.key_expiry_webhook_secret_rotated_at	str	''	ISO-8601 timestamp when key_expiry_webhook_secret_previous was last set by rotation (#1680).
alerts.key_expiry_warn_days	int	14	Days before expires_at for doctor warnings, Prometheus expiring_soon gauge, and key_expiry webhooks (#1546). Env: DAARI_ALERTS__KEY_EXPIRY_WARN_DAYS.
skills_system_prefix	str	''
model_groups	dict	{}	Named model groups (exact names or globs such as claude-*). Keys and teams reference them by name; enforcement is the union of allowed_models and the referenced groups, intersected across team and key.

Per-key and per-team rpd (requests per UTC day, 0 = unlimited) is not a rate_limit.* setting. Set it on the key or team (daari keys create/update --rpd, daari keys team-create/update --rpd). See auth and keys.

Per-key modality-family RPM/TPM (chat / embeddings / images / audio / moderations / rerank / ocr / mcp / systemone / other) is also not a rate_limit.* setting — store it on the virtual key as metadata rate_families (see auth-and-keys). Unset families keep global key rpm/tpm only (#1099).