.env.example

# Server encryption key for API key storage (generate with: node -e "console.log(require('crypto').randomBytes(32).toString('hex'))").
# Required in production. Outside production it is optional: when unset, a key is
# auto-generated and written to a file named .encryption-key next to the SQLite
# database (not inside it), with 0600 permissions. Precedence: this env var, then
# that key file, then a legacy key found in the old settings table (migrated to
# the file on first boot), then a freshly generated key.
ENCRYPTION_KEY=your-64-char-hex-key-here

# Server port (default: 3001)
PORT=3001

# Network interface the server listens on. Default '::' (dual-stack IPv4+IPv6);
# hosts with IPv6 disabled fall back to IPv4 automatically. Set 0.0.0.0 for
# IPv4-only or 127.0.0.1 to restrict to localhost. (Different from HOST_BIND
# below, which only affects Docker port publishing.)
# HOST=::

# First-run setup: the first account is created through the dashboard. A browser
# on the same machine as the server can do this with no extra step. If the server
# is reachable from other devices, creating that first account also requires a
# one-time setup code, printed in the server logs at startup while no account
# exists. This stops a stranger from claiming a freshly exposed install.

# Docker only: which host interface the container's port is published on.
# Defaults to 127.0.0.1, so the dashboard/API is reachable only from the
# machine running Docker. To open it to other devices on your LAN (e.g. a
# Raspberry Pi reached at http://192.168.1.x:3001), set HOST_BIND=0.0.0.0.
# Only do this on a trusted network — the proxy is single-user and guarded
# only by the unified API key.
# HOST_BIND=0.0.0.0

# Max /v1 proxy requests per minute per client IP (default: 120). Set to 0
# to disable proxy rate limiting.
# PROXY_RATE_LIMIT_RPM=120

# SSE heartbeat interval in ms for streaming requests (#1642). Reasoning
# models can go minutes without a first token, and clients with an idle
# watchdog (Claude Code) abort during that silence. Set to e.g. 15000 and
# each waiting stream emits a content-free SSE ping every 15s so the client
# keeps waiting. OFF by default: a ping commits the response, so a stream
# that dies after one ends with an error frame instead of failing over to
# the next provider invisibly. Floor 1000 ms.
# STREAM_KEEPALIVE_MS=15000

# Max /api dashboard requests per minute per client IP (default: 600). This is a
# flood guard — login has its own per-email lockout and key export its own much
# tighter cap. Set to 0 to disable.
# ADMIN_RATE_LIMIT_RPM=600

# JSON body limit in MB for the inference surfaces (/v1, /v1beta, /mcp,
# Ollama /api/*) — default: 25. Vision requests embed base64 images in the
# body and replay them with every turn, so large sessions can clear 10MB;
# this ceiling only gates parsing — inbound image normalization (below)
# shrinks the payload afterwards. Raise it if bigger payloads are rejected
# with 413 'request_too_large'.
# REQUEST_BODY_LIMIT_MB=25

# ChatGPT / MCP inference. ask_freellmapi uses "auto" unless this is set to
# a model id or named chain (for example auto:coding). The timeout applies only
# to the MCP inference tool; normal /v1 client requests keep their own limits.
# MCP_INFERENCE_DEFAULT_MODEL=auto
# MCP_INFERENCE_TIMEOUT_MS=120000

# Inbound image normalization: data-URL images over the threshold are
# downscaled to the long-edge cap and re-encoded (JPEG, or PNG only when the
# alpha channel carries real transparency) before routing, shrinking replayed
# screenshots ~6-10x. Also normalizes exotic formats upstreams reject
# (bmp/gif/tiff/avif/webp → jpeg/png); gif keeps its first frame only. Every
# upstream resizes internally anyway (OpenAI to 2048px, Anthropic to 1568px),
# so pixels beyond the cap are transport waste. Set IMAGE_NORMALIZE=off to
# disable entirely.
# IMAGE_NORMALIZE=on
# IMAGE_NORMALIZE_MAX_DIMENSION=2048
# IMAGE_NORMALIZE_THRESHOLD_KB=1024
# IMAGE_NORMALIZE_QUALITY=90

# Outbound proxy for provider requests. Normally set in the dashboard
# (Keys → Outbound proxy); these env vars are for headless installs.
# Schemes: http, https, socks4, socks4a, socks5, socks5h. The `h`/`a` variants
# resolve DNS at the proxy, which is what you want on a DNS-poisoned network.
# Precedence: PROXY_URL → dashboard setting → ALL_PROXY → HTTPS_PROXY →
# HTTP_PROXY. The standard vars are also read in their lower-case spellings.
# PROXY_URL=socks5h://127.0.0.1:1080
# ALL_PROXY=socks5h://127.0.0.1:1080
# HTTPS_PROXY=http://proxy.corp.com:8080
# HTTP_PROXY=http://proxy.corp.com:8080
#
# Application-layer Fetch Relay (opt-in). The relay target and its separate
# bearer credential use Fetch-Relay-Target and Fetch-Relay-Authorization.
# See docs/en/proxy/01-fetch-relay.md.
# PROXY_MODE=fetch-relay
# PROXY_URL=https://relay.example.workers.dev
# FETCH_RELAY_TOKEN=generate-a-long-random-token
#
# IN DOCKER, 127.0.0.1 IS THE CONTAINER, NOT YOUR MACHINE. If the proxy runs
# on the host (Clash, v2rayN, sing-box, a corporate client), use the host's
# address instead, and make sure the proxy listens on more than loopback
# (Clash: `allow-lan: true`):
# PROXY_URL=socks5h://host.docker.internal:7890
# The bundled docker-compose.yml maps host.docker.internal to the host gateway
# so this works on plain Linux Docker too, not just Docker Desktop.

# Hosts that must be reached directly, bypassing the proxy above. Comma-
# separated; a bare domain also covers its subdomains, and `*` disables the
# proxy entirely.
# NO_PROXY=localhost,127.0.0.1,.internal.corp

# Optional fail-closed outbound policy. These platform IDs MUST use a proxy:
# if the proxy is missing, disabled, bypassed, or its agent cannot be built,
# return an explicit error instead of silently connecting directly.
# Non-listed platforms keep the normal routing behavior. This controls NETWORK
# transport only; it does not disable model/provider failover in auto routing.
# Example: require Gemini to use your configured SOCKS5 tunnel.
# PROXY_REQUIRED_PLATFORMS=google

# Local and LAN destinations (localhost, 127.0.0.0/8, ::1, 0.0.0.0, RFC1918 /
# ULA / CGNAT addresses) always bypass the proxy — a remote proxy has no route
# back to your own machine, and an IP literal makes Tor log "giving Tor only an
# IP address" and refuse the connection. That is the right default for a local
# Ollama / llama.cpp / LM Studio endpoint.
# Set this to true only if proxying them is the point — e.g. an `ssh -D`
# dynamic tunnel, where http://127.0.0.1:11434 through the SOCKS proxy is meant
# to reach the REMOTE host's Ollama.
# FREEAPI_PROXY_LOCAL_DESTINATIONS=true

# Per-provider daily request cap, named PROVIDER_DAILY_REQUEST_CAP_<PLATFORM>
# (platform name upper-cased). Overrides the built-in default for that provider;
# set to 0 to disable the cap. Example — cap OpenRouter at 50 requests/day:
# PROVIDER_DAILY_REQUEST_CAP_OPENROUTER=50

# Per-provider chat HTTP timeout in milliseconds, named
# PROVIDER_TIMEOUT_<PLATFORM> (platform name upper-cased). Overrides the
# built-in default for that provider (60s for most OpenAI-compatible platforms,
# 180s NVIDIA, 120s Ollama Cloud / AI Horde / custom providers, 60s Google as
# registered). On streaming requests this value also serves as the FIRST-BYTE
# grace budget: how long the stream may stay silent before its first byte
# (issue #584 — providers like NVIDIA NIM send SSE headers instantly, then
# prefill long prompts for minutes). Set to 0 to disable the timeout entirely.
# Read at startup. Example — give NVIDIA NIM's slow reasoning models 5 minutes:
# PROVIDER_TIMEOUT_NVIDIA=300000

# ModelScope (魔搭) validates API keys with a paid 1-token chat completion
# (GET /v1/models does not enforce auth), which spends the account's
# magic-grain (魔粒) quota per probe — observed as 2 魔粒 per ultra-tier
# request. To stop the 5-minute health pass from burning ~288 paid probes
# per key per day, a successful validation is cached per key for this long
# (milliseconds, default 86400000 = 24h). Set to 0 to probe on every pass.
# A revoked key is still caught by the next real request's 401.
# MODELSCOPE_VALIDATE_CACHE_MS=86400000

# Mid-stream inactivity timeout for streaming responses, in milliseconds
# (default: 90000). If an SSE stream goes this long without a single byte
# AFTER the first byte arrived, the gateway gives up on it (before any output
# reached the client this fails over to the next candidate; after output was
# sent the stream ends with an error frame). Time-to-FIRST-byte is governed by
# the chat timeout above instead (whichever of the two is larger). Providers
# that queue heavily on their free tiers can pause longer than 90s mid-stream —
# raise this if streamed responses stop early. Set to 0 to disable the stall
# watchdog entirely.
# PROVIDER_STREAM_STALL_TIMEOUT_MS=90000
#
# Per-provider variant, named PROVIDER_STREAM_STALL_TIMEOUT_<PLATFORM>
# (platform name upper-cased). Precedence: per-platform > global (above) >
# built-in 90s default. Example — let Ollama Cloud pause 4 minutes mid-stream
# without loosening the watchdog everywhere else:
# PROVIDER_STREAM_STALL_TIMEOUT_OLLAMA=240000

# Base cadence for the scheduled key health pass, in milliseconds (default:
# 300000 = 5m, minimum 60000; lower or invalid values fall back to the default).
# Each pass applies ±20% jitter. A faster cadence sends more validation traffic,
# which costs quota where a provider's key check is metered. Read at startup.
# HEALTH_CHECK_INTERVAL_MS=300000

# Optional destination for soft budget alerts: when a provider key crosses 80%
# or 100% of its monthly request/token cap, the alert is POSTed to this URL as
# JSON (type: freellmapi.key_budget_threshold) in addition to the Logs page
# event. http(s) only; delivery is fire-and-forget with a 10s timeout. Works
# with Slack/Discord via a workflow webhook. Read when an alert fires.
# BUDGET_ALERT_WEBHOOK_URL=https://hooks.example.test/abc

# Opt-in background model probes (#1438): every 30 minutes, send a tiny
# "Reply with OK" chat to up to MODEL_PROBE_MAX_PER_PASS (default 3) model+key
# pairs and record whether each model answers. Off by default because every
# probe spends free-tier quota. POST /api/health/probe-models runs one pass on
# demand either way.
# MODEL_PROBE_ENABLED=1
# MODEL_PROBE_MAX_PER_PASS=3

# Base wall-clock budget for provider failover, in milliseconds (default: 45000).
# The initial attempt and one failover hop are always allowed. After that,
# exhaustion stops new attempts and can cancel an in-flight attempt still
# waiting for its first usable response. Streaming responses disarm this timer
# when the first usable response arrives. Successful TTFB history can widen
# the budget for a slow endpoint (see below). Set to 0 to disable the budget
# entirely, including adaptive widening, and retry until the attempt cap.
# Can also be set at runtime via the fallback_time_budget_ms settings key,
# which takes precedence.
# FALLBACK_TIME_BUDGET_MS=45000

# Adaptive patience for slow endpoints: successful time-to-first-byte (TTFB)
# samples are grouped by platform + endpoint_scope. The effective budget is
# max(base budget, historical P95 TTFB + this buffer), in milliseconds.
# Endpoints with no samples, and those whose P95 + buffer fits the base budget,
# keep the base budget. The settings key slow_endpoint_buffer_ms takes
# precedence over this env var. Set the buffer to 0 for no extra padding.
# SLOW_ENDPOINT_BUFFER_MS=10000
# Successful TTFB sample window, in milliseconds (default: 7 days).
# The settings key ttfb_budget_window_ms takes precedence.
# TTFB_BUDGET_WINDOW_MS=604800000
# Sample-weight half-life, in milliseconds (default: 2 days). Older successes
# count less so the budget adapts as an endpoint speeds up or slows down.
# The settings key ttfb_budget_half_life_ms takes precedence.
# TTFB_BUDGET_HALF_LIFE_MS=172800000
# Successful samples an endpoint needs before its history widens the budget
# (default: 5). The widened budget is capped at three times the base budget.
# The settings key ttfb_budget_min_samples takes precedence.
# TTFB_BUDGET_MIN_SAMPLES=5
# Set to 1 to disable adaptive TTFB budgets and use only the base budget.
# TTFB_BUDGET_DISABLED=1

# Add X-Fallback-Detail to responses (default: off). X-Fallback-Trail already
# names every hop that failed and why; the detail header adds what each one
# COST — "groq/llama-3.3-70b key1=rate_limited t=0+39000ms msg=..." — so a
# caller can tell one 39s stall apart from four fast failures without opening
# the dashboard. Off by default because it puts hop timings and provider error
# text on the response. Can also be set at runtime via the
# expose_fallback_detail_header settings key, which takes precedence.
# FALLBACK_DETAIL_HEADER=1

# Check tool-call arguments against the schema the caller declared, and fail
# over to another model when they violate it (default: off). lib/tool-args.ts
# already repairs what it can prove is repairable; this catches what is left,
# which today reaches Anthropic clients as a tool_use block with `input: {}` and
# no indication anything went wrong. A different model usually gets the same
# call right. Off by default because a false positive costs a failover hop —
# turn it on, watch X-Fallback-Trail for `invalid_tool_arguments`, then decide.
# Applies wherever the turn can still fail over — every non-streaming surface,
# and the streams that buffer tool calls to the end (/v1/chat/completions,
# /v1/messages, the Gemini and Ollama emulations). /v1/responses streaming is
# the exception: it commits on the first tool-call delta, before the arguments
# are complete, so a verdict there could only break a stream in flight.
# Can also be set at runtime via the validate_tool_arguments settings key,
# which takes precedence.
# VALIDATE_TOOL_ARGUMENTS=1

# Request guardrails, both OFF (0) by default. Also runtime-tunable via the
# request_max_tokens_budget / max_consecutive_upstream_fails settings keys
# (dashboard API PUT /api/settings/guardrails), which take precedence.
#
# Per-request token budget: estimated input + requested max_tokens must fit
# this ceiling or the request is rejected with a 413 before any provider is
# tried; a request that sent no max_tokens gets its output capped to the
# remainder instead.
# REQUEST_MAX_TOKENS_BUDGET=0
#
# Failover circuit breaker: after this many consecutive upstream failures in
# one request, stop the failover chain with a 503 instead of trying every
# remaining candidate of an unhealthy pool.
# MAX_CONSECUTIVE_UPSTREAM_FAILS=0

# Opt-in response cache. When on, a successful /v1/chat/completions answer --
# streaming or not -- is stored in a bounded in-memory LRU keyed by a hash of the
# request, so an IDENTICAL later request is served from memory without spending
# any provider quota. Exact match only (no fuzzy matching), so a near-miss never
# returns a different prompt's answer. Off by default; set to true to enable.
# Can also be toggled at runtime via the dashboard (PUT /api/cache/config), and
# per-request with the `X-FreeLLM-Cache: on|off` header.
# RESPONSE_CACHE=false
# Persist cached non-streaming answers to SQLite so they survive a restart
# (streamed replays are memory-only and never persisted) (the daily
# quota-reset re-run pattern). Only applies when the cache above is on. Note
# this writes PLAINTEXT model responses to the database file, which the
# memory-only cache never did; set it to false to keep caching but hold
# everything in memory, as before. Entries are removed from disk the moment
# they leave the in-memory cache (TTL expiry, eviction, DELETE /api/cache).
# RESPONSE_CACHE_PERSIST=true
# How long a cached answer stays fresh, in seconds (default: 3600 = 1 hour).
# RESPONSE_CACHE_TTL_SECONDS=3600
# Only cache requests at/below this temperature; higher temperatures want fresh
# variety, so they are never cached. Default 1.0 caches everything when enabled;
# lower it (e.g. 0.2) to cache only near-deterministic calls.
# RESPONSE_CACHE_MAX_TEMPERATURE=1.0
# Hard cap on stored entries, shared by the JSON and streaming stores; the
# least-recently-used are evicted past this.
# RESPONSE_CACHE_MAX_ENTRIES=5000

# Request-side prompt/context compression. The default is off. `lossless`
# applies whitespace hygiene, exact repeated-block references, and reversible
# tabular-JSON encoding; `standard` also filters tool output and superseded file
# reads; `aggressive` adds age/relevance/budget condensation. The dashboard can
# change this live under Settings. A request may lower the configured mode with
# `X-FreeLLM-Compress`, but cannot override a global `off` master switch.
# FREELLMAPI_COMPRESSION=off

# Custom-provider URL policy. Cloud metadata and link-local addresses
# (169.254.169.254, metadata.google.internal, fe80::/10, ...) are ALWAYS
# blocked as custom provider base URLs. By default localhost and private LAN
# addresses stay allowed so local Ollama / LM Studio / llama.cpp setups work.
# If you host FreeLLMAPI on a VPS or anywhere the dashboard is reachable by
# others, set this to also block loopback and RFC1918/ULA private ranges.
# FREEAPI_BLOCK_PRIVATE_PROVIDER_URLS=true

# Request analytics retention. Set either value to 0 to disable that limit.
REQUEST_ANALYTICS_RETENTION_DAYS=90
REQUEST_ANALYTICS_MAX_ROWS=100000

# Provider quota observation log (the audit trail behind the per-key quota
# state the dashboard shows). Only the newest row per pool is ever read, so the
# history is pruned daily by age and by count. 0 disables that half.
# QUOTA_OBSERVATIONS_RETENTION_DAYS=7
# QUOTA_OBSERVATIONS_MAX_ROWS=50000

# Persisted server logs behind the dashboard's log viewer. Only warn/error
# lines are written to the database (the live view is an in-memory ring), so
# these bounds are far tighter than the analytics ones above. Either value at 0
# disables that limit.
# SERVER_LOGS_RETENTION_DAYS=7
# SERVER_LOGS_MAX_ROWS=50000

# Web-search attempt log behind the dashboard's Search page (search_requests).
# Every search attempt appends a row (an agent web_search can issue up to 20),
# so the log is bounded by age and count on the same pass as the analytics
# prune. 0 disables that half.
# SEARCH_LOGS_RETENTION_DAYS=90
# SEARCH_LOGS_MAX_ROWS=100000

# Decision-model attempt log behind the dashboard's Models → Decision page
# (decision_requests). Every /v1/decisions attempt appends a row, so the log
# is bounded by age and count on the same pass as the analytics prune.
# 0 disables that half.
# DECISION_LOGS_RETENTION_DAYS=90
# DECISION_LOGS_MAX_ROWS=100000

# Per-request caller identity (client IP + User-Agent) recorded into request
# analytics and shown in the dashboard's "Recent calls" table. Set to false
# to store nulls instead (aggregate analytics are unaffected).
# REQUEST_ANALYTICS_LOG_CLIENT=true

# Optional SQLite location. Useful on hosts where only one directory is mounted
# persistently, or when you want to keep the DB outside server/data.
# FREEAPI_DB_PATH=/app/server/data/freellmapi.db

# Whether to restrict the directory holding the database to this account, so the
# SQLite -wal/-shm sidecars are protected too (they are created on first write,
# after any startup permission check could reach them). Done automatically for
# the default server/data directory and for any directory FreeLLMAPI creates
# itself; skipped for a FREEAPI_DB_PATH pointing somewhere pre-existing, since
# that may be a shared location this server should not be locking down. Set to 1
# if the directory belongs to this account alone, or 0 to disable entirely.
# FREEAPI_DB_DIR_HARDENING=1

# Optional encrypted SQLite backup. On startup, FreeLLMAPI restores this backup
# if the configured DB file is missing; while running, it uploads a fresh backup
# periodically. File paths and HTTP(S) URLs are supported.
# FREEAPI_DB_BACKUP_PATH=/app/server/data/freellmapi.db.backup
# FREEAPI_DB_BACKUP_URL=https://example.com/freellmapi.db.backup
# FREEAPI_DB_BACKUP_TOKEN=
# FREEAPI_DB_BACKUP_KEY=use-a-separate-64-char-hex-key-or-omit-to-use-ENCRYPTION_KEY
# FREEAPI_DB_BACKUP_INTERVAL_MS=300000

# Automatic dump files on the Backups page (#1675): how many dumps created by
# the scheduler or the desktop updater's pre-update snapshot to keep. Older
# ones are deleted (file + row) on the next dump write. Manual and pre-restore
# dumps are never auto-deleted. 0 keeps every auto dump (disables the prune).
# AUTO_BACKUP_DUMP_RETENTION=10

# Optional declarative startup config. Use either inline JSON or a path to a JSON
# file. It is applied idempotently after migrations on every boot — before the
# HTTP listener starts, so an "admin" entry closes the first-run setup window
# entirely. Optional "admin" creates the first dashboard account (only while no
# account exists); optional "license" activates a freellmapi.co Premium key.
# The JSON carries plaintext secrets — protect it like ENCRYPTION_KEY.
# FREEAPI_CONFIG_PATH=/app/server/data/freellmapi.config.json
# FREEAPI_CONFIG_JSON={"admin":{"email":"ops@example.com","password":"change-me"},"license":"...","keys":[{"platform":"groq","key":"gsk_...","label":"main"}],"routing":{"strategy":"balanced"}}

# Context handoff on model switch. When enabled, FreeLLMAPI injects a compact
# system message into the outbound request whenever a session switches from one
# model to another (e.g. after a fallback). Off by default.
# FREELLMAPI_CONTEXT_HANDOFF=on_model_switch

# Manual application update checker. Set to off to hide it from Settings and
# prevent Git discovery or outbound update-check requests. Enabled by default.
# This also switches off the automatic release reminder, which is a separate
# dashboard setting (Settings > General) and is off until turned on there.
# FREELLMAPI_UPDATE_CHECK=off
# Optional token used only for update checks against GitHub. The checker is
# anonymous by default; use a narrowly scoped token only if higher rate limits
# are needed. Generic GITHUB_TOKEN values are intentionally ignored.
# FREELLMAPI_UPDATE_GITHUB_TOKEN=
# FREELLMAPI_COMMIT_SHA and FREELLMAPI_INSTALL_METHOD are injected by official
# Docker and desktop builds and normally should not be set in .env.
# Docker "Update now" (Settings > Check for updates). Set a long random token,
# e.g. `openssl rand -hex 32`, then start with
# `docker compose --profile autoupdate up -d`. See docker-compose.yml.
# WATCHTOWER_TOKEN=

# Comma-separated extra origins allowed to call the API from a browser.
# localhost:5173, 127.0.0.1:5173, [::1]:5173 are allowed by default for the
# Vite dev dashboard. Only set this if you serve the dashboard from a
# different host than the API (e.g. http://my-server.local).
# DASHBOARD_ORIGINS=http://my-server.local,http://192.168.1.50

# Content-Security-Policy `upgrade-insecure-requests` directive (#682).
# Auto (unset): emit only when the request arrives over TLS or behind an
# HTTPS reverse proxy (X-Forwarded-Proto: https). Plain-HTTP LAN installs
# (HOST_BIND=0.0.0.0, no TLS in front) stay renderable.
# true:  force the directive on even over plain HTTP (rarely useful).
# false: force it off even behind HTTPS (e.g. mixed-content reverse proxies).
# CSP_UPGRADE_INSECURE_REQUESTS=false

# Trust forwarded headers (X-Forwarded-For / X-Forwarded-Proto) from a reverse
# proxy so Analytics and IP-based rate limiting see the real client (#1024).
# Default (unset or false): do NOT trust forwarded headers — direct callers
# cannot spoof the client IP.
# true: trust every proxy hop (only use on a fully trusted network).
# A number: trust that many hops in front of the server (TRUST_PROXY=1 for a
# single reverse proxy such as Caddy, nginx or Traefik on the same host).
# Otherwise: a comma-separated list of proxy addresses/CIDRs to trust, e.g.
#   TRUST_PROXY=100.64.0.0/10,192.168.1.10
# TRUST_PROXY=false

# Host-header allowlist against DNS rebinding (#1637). Off by default. Once
# set, a request whose Host header names anything other than loopback
# (localhost, 127.0.0.1, ::1, *.localhost) or a listed hostname gets 403
# before auth or routing. List every hostname/IP you reach the server by,
# comma-separated. Loopback-only:
#   ALLOWED_HOSTS=localhost
# LAN or reverse proxy:
#   ALLOWED_HOSTS=192.168.1.50,freellmapi.home.arpa
# Unset (or *) leaves the check off.
# ALLOWED_HOSTS=

# Per-model routing score multipliers (#738), JSON object of model id → 0..2.
# 1.0 leaves a model unchanged, below 1 demotes it, above 1 promotes it, and 0
# means the automatic router never picks it (Manual priority still can).
# Matching is by model id alone, across all platforms serving that id, and the
# match is exact and case-sensitive. Copy ids from the dashboard Models page:
# an id that matches nothing is silently ignored (the boot log flags it once the
# catalog has synced).
# MODEL_ROUTING_OVERRIDES={"gpt-4o": 0.2, "gpt-oss-120b": 1.5}

# --- Advanced / embedding ---------------------------------------------------
# Path to a prebuilt client `dist` directory to serve. Defaults to the bundled
# client build; set this only if you build the dashboard separately.
# CLIENT_DIST=/path/to/client/dist

# Explicit path to the .env file to load. Useful for embedders (e.g. the
# desktop app, where the code runs from inside a bundle). Defaults to ./.env.
# FREEAPI_ENV_PATH=/path/to/.env