NODE_ENV=development # Compose Postgres role (source-checkout and images stacks). Generate with: # openssl rand -hex 16 # Use a URI-safe value (hex from openssl is fine). Special characters such as # @ : / ? # % break DATABASE_URL interpolation inside Compose. # Put the same value in DATABASE_URL when using the optional host publish overlay. # Existing pgdata volumes keep user/password/db from first init; keep those # values in .env, or ALTER ROLE / rename in place. Recreate the volume only # after a backup (or when disposable); down -v deletes all Postgres state. POSTGRES_USER=rakazo POSTGRES_PASSWORD= POSTGRES_DB=rakazo # The app's own login (rakazo_app): no superuser, no schema changes, and subject # to row-level security. `pnpm db:migrate` creates it with this password. # Generate with: openssl rand -hex 16 POSTGRES_APP_PASSWORD= # Optional host-publish overlay port. Keep the DATABASE_*_URL ports in sync. POSTGRES_HOST_PORT=5433 # Host-side clients only (pnpm db:migrate / pnpm dev). Requires the postgres-host # overlay so 127.0.0.1:5433 is published. Inside Compose, api/worker ignore these # and connect to the postgres service with POSTGRES_* above. # The app connects as the runtime login; migrations run as the owner. DATABASE_URL=postgres://rakazo_app:REPLACE_WITH_POSTGRES_APP_PASSWORD@127.0.0.1:5433/rakazo DATABASE_MIGRATION_URL=postgres://rakazo:REPLACE_WITH_POSTGRES_PASSWORD@127.0.0.1:5433/rakazo BETTER_AUTH_SECRET=replace-with-32-plus-character-secret # Local loopback: localhost and 127.0.0.1 are interchangeable browser hosts. BETTER_AUTH_URL=http://127.0.0.1:5173 API_URL=http://127.0.0.1:3100 # Listener address. Keep loopback unless a container or trusted reverse proxy must reach the API. API_HOST=127.0.0.1 # Reverse proxies whose forwarded client address the API believes: IPs, CIDR ranges, or hostnames. # Empty trusts loopback; in a container, also the networks it is attached to (local-networks). # none trusts no proxy. TRUSTED_PROXIES= # The header those proxies set. Default x-forwarded-for. CLIENT_IP_HEADER= WEB_ORIGIN=http://127.0.0.1:5173 # Optional second address for the same install (tailnet or LAN), comma-separated. https:// or # http:// on loopback/private IPs only. Trusted for sign-in and CORS; see docs/self-host-configuration.md. RAKAZO_EXTRA_AUTH_ORIGINS= # pnpm dev only: 1 serves the web app to your Tailscale devices over HTTPS (README). RAKAZO_DEV_TAILNET= # pnpm dev only: a command that adds secrets to the environment and then runs the one after it, # space-separated, such as a secrets manager's run command ending in `--` (README). RAKAZO_DEV_RUN_WITH= # Optional Compose web bind address; use 0.0.0.0 for LAN access (docs/self-host-clients.md). RAKAZO_WEB_BIND=127.0.0.1 # Optional https:// origins added to the web Content-Security-Policy, space-separated, for a # script your CDN injects (for example analytics). See docs/self-host-clients.md#content-security-policy. # RAKAZO_CSP_SCRIPT_SRC=https://static.example.com # RAKAZO_CSP_CONNECT_SRC=https://beacon.example.com # Optional web colors and fonts: a theme file the web service can read (docs/self-host-clients.md#colors-and-fonts). # RAKAZO_THEME=/theme/theme.json # Optional: refuse clients built before these UTC times (YYYYMMDDHHMMSS) so they update. # See docs/self-host-clients.md#client-updates. # RAKAZO_MIN_CLIENT_BUILD=web=20261022093015,mobile=20261001000000 # Optional Crisp chat: CRISP_WEBSITE_ID on the paywall (API and web), PUBLIC_ on the website. # CRISP_WEBSITE_ID= # PUBLIC_CRISP_WEBSITE_ID= PUBLIC_POSTHOG_KEY= PUBLIC_POSTHOG_HOST=https://us.i.posthog.com SIGNUPS_ENABLED=true # Opt-in: reject new passwords found in Have I Been Pwned (sends a 5-character hash prefix). AUTH_BREACHED_PASSWORD_CHECK_ENABLED=false # Optional Google sign-in. Set both, or leave both empty. See docs/self-host-accounts.md. GOOGLE_CLIENT_ID= GOOGLE_CLIENT_SECRET= # Optional Apple sign-in. The API signs the client secret from this .p8 key; do not paste a JWT. # Put real newlines in APPLE_PRIVATE_KEY, or write \n for each line break. Set the file path # instead when the key is mounted into the API container. See docs/self-host-accounts.md. APPLE_CLIENT_ID= APPLE_TEAM_ID= APPLE_KEY_ID= APPLE_PRIVATE_KEY= # APPLE_PRIVATE_KEY_FILE= # Comma-separated emails or @domains. Applied on every API start; empty allows any email. SIGNUP_ALLOWLIST= # Optional: lets your identity provider add and remove accounts over SCIM (32+ random characters). # See docs/self-host-accounts.md#directory-provisioning-scim. SCIM_BEARER_TOKEN= # Optional: native apps allowed to use this domain's passkeys. See docs/self-host-accounts.md#passkeys. PASSKEY_APPLE_APP_IDS= PASSKEY_ANDROID_CERT_SHA256= ENCRYPTION_KEY=replace-with-64-random-hex-characters # Optional content master keys, active first: id:key,older-id:key. See docs/content-encryption.md. ENCRYPTION_MASTER_KEYS= # Optional: aws-kms wraps content keys with a key that never leaves AWS KMS. See docs/content-encryption.md. MASTER_KEY_PROVIDER=local AWS_KMS_KEY_ARN= AWS_KMS_ACCESS_KEY_ID= AWS_KMS_SECRET_ACCESS_KEY= # on|off: encrypt new content at rest. Empty is off in development and on elsewhere. CONTENT_ENCRYPTION= # Optional credential storage; API and worker must share this scope. See docs/infisical-secrets.md. SECRET_STORE=encrypted INFISICAL_URL= INFISICAL_ALLOW_INSECURE_HTTP=false INFISICAL_CLIENT_ID= INFISICAL_CLIENT_SECRET= INFISICAL_PROJECT_ID= INFISICAL_ENVIRONMENT= INFISICAL_FOLDER= DATA_DIR=./data # Optional: send the browser on bot computers through an upstream proxy (Railway computers). # http://user:pass@host:port or socks5://user:pass@host:port. Each computer runs tinyproxy on # loopback with the credentials, and a managed Chrome policy points the browser at it. # Only the browser uses it; terminal commands still go direct. Bots run as root, so anyone who can # use a bot can read these credentials: use a dedicated, usage-capped proxy login. # RAKAZO_COMPUTER_PROXY= # Optional: keep computer homes and artifacts in S3-compatible or Azure Blob storage instead of # DATA_DIR. Set the bucket (the container for Azure Blob) to turn it on; the protocol is s3 (default) # or azure-blob. See docs/self-host-storage.md#shared-storage-migration. OBJECT_STORAGE_BUCKET= OBJECT_STORAGE_PROTOCOL= OBJECT_STORAGE_ENDPOINT= OBJECT_STORAGE_REGION= OBJECT_STORAGE_ACCESS_KEY_ID= OBJECT_STORAGE_SECRET_ACCESS_KEY= OBJECT_STORAGE_PREFIX= OBJECT_STORAGE_PATH_STYLE= SANDBOX_SUPERVISOR_URL=http://127.0.0.1:7091 # Dedicated credentials (distinct from BETTER_AUTH_SECRET). Required for Docker sandboxes / # screen proxy outside local development. SANDBOX_SUPERVISOR_TOKEN=replace-with-32-plus-character-supervisor-token # Per-bot-computer ceilings. Defaults 2g / 2 CPUs / 2048 pids; "0", "none" or "unlimited" removes # one. Memory also caps swap, and must be at least 6m (Docker's own minimum). RAKAZO_COMPUTER_MEMORY= RAKAZO_COMPUTER_CPUS= RAKAZO_COMPUTER_PIDS_LIMIT= SCREEN_PROXY_SECRET=replace-with-32-plus-character-screen-proxy-secret SANDBOX_PROVIDER=docker # true lets Settings → Computer choose a provider instead of SANDBOX_PROVIDER. # COMPUTER_SETTINGS_ENABLED=false # none | emulator | cursor | coder. Live agents require a key and an explicitly authorized Space. CLOUD_AGENT_PROVIDER=none CURSOR_API_KEY= # coder: the workspace URL (https://coder.example.com/@owner/workspace) and a Coder token. CLOUD_AGENT_URL= CLOUD_AGENT_API_KEY= # Grant this Space access to the repositories the cloud agent credential can reach. CLOUD_AGENT_SPACE_ID= # none | link | emulator. Lets bots request purchases the user approves in their wallet. # PAYMENT_WALLET_PROVIDER=none # Optional registered OAuth client id; defaults to the provider's public client. # PAYMENT_WALLET_CLIENT_ID= # Optional Docker daemon socket. Defaults to /var/run/docker.sock (Unix) or # //./pipe/docker_engine (Windows). Set for rootless Docker, Colima, or custom paths. # DOCKER_HOST takes precedence over DOCKER_SOCKET when both are set. # DOCKER_SOCKET= # Enable for a host-run supervisor using a local Docker Desktop daemon: control # uses a random token-guarded 127.0.0.1 port because container IPs are unreachable. # Leave unset for Compose, remote Docker daemons, or native Linux Docker Engine. # Changing this setting replaces incompatible computers on their next provision; # their mounted homes are preserved. # SANDBOX_CONTROL_VIA_LOOPBACK=true # Novita: build the desktop template with pnpm sandbox:novita-template before use. NOVITA_API_KEY= # Optional template name for Novita; defaults to rakazo-computer (2 CPU, 4 GB). SANDBOX_TEMPLATE= DAYTONA_API_KEY= DAYTONA_API_URL= DAYTONA_TARGET= # Optional snapshot name for bot computers; unset uses the Daytona server default. DAYTONA_SNAPSHOT= CREATEOS_SANDBOX_API_KEY= # Optional CreateOS API base URL; defaults to https://api.sb.createos.sh. CREATEOS_SANDBOX_BASE_URL= # Optional CreateOS sandbox shape/rootfs; defaults to s-2vcpu-2gb and desktop:1. CREATEOS_SANDBOX_SHAPE= CREATEOS_SANDBOX_ROOTFS=desktop:1 # Railway workspace/account API token; keep it on the API and worker only. RAILWAY_MAX_WORKSPACE_FILE_BYTES=67108864 RAILWAY_API_TOKEN= # Railway supplies this automatically when hosted there. RAILWAY_ENVIRONMENT_ID= BOAT_API_KEY= # Optional Boat API base URL; defaults to https://boat.dev/api/v1. BOAT_API_URL= BOXD_API_KEY= # Optional boxd organization for computers; defaults to the key owner's own machines. BOXD_ORG= # instabox computers (SANDBOX_PROVIDER=instabox) use the key's organization. INSTABOX_API_KEY= # Wuying cloud computer (SANDBOX_PROVIDER=wuying). All bots, users and spaces, including # Private Computers, share its Windows desktop and account permissions. Trusted deployments only. # Keep the AccessKey on the API and worker only; the browser gets a # one-time authCode per screen connection. WUYING_ACCESS_KEY_ID= WUYING_ACCESS_KEY_SECRET= # Optional region; defaults to cn-hangzhou. WUYING_REGION_ID= WUYING_DESKTOP_ID= # End user the desktop is assigned to; authCodes are issued for this user. WUYING_END_USER_ID= # MCP gateway bound to the desktop (commands, files, screen and input). WUYING_MCP_URL= WUYING_MCP_TOKEN= # Incus app containers (SANDBOX_PROVIDER=incus); see docs/self-host-sandbox-providers.md#incus. # The Incus API address and a client certificate trusted for one restricted project. PEM values # may be one line with \n escapes. Keep the key on the API and worker only. INCUS_SANDBOX_URL= INCUS_SANDBOX_CLIENT_CERT= INCUS_SANDBOX_CLIENT_KEY= # Optional: the server's own certificate, trusted exactly. Leave unset when the server presents a # certificate from a public CA for INCUS_SANDBOX_URL's host name. INCUS_SANDBOX_SERVER_CERT= # Optional; default is the default project. INCUS_SANDBOX_PROJECT= # Optional; defaults to ghcr.io/elie222/rakazo/computer:edge. Pin a release tag or digest. INCUS_SANDBOX_COMPUTER_IMAGE= # Optional per-computer limits in Incus units; defaults 2, 2GiB and 2048. INCUS_SANDBOX_CPU_LIMIT= INCUS_SANDBOX_MEMORY_LIMIT= INCUS_SANDBOX_PROCESS_LIMIT= AGENT_RUNTIME=pi # Optional Pi traces via OTLP/HTTP; empty endpoint means no SDK initialization. OTEL_EXPORTER_OTLP_ENDPOINT= # Standard comma-separated key=value headers; keep credentials private. OTEL_EXPORTER_OTLP_HEADERS= OTEL_SERVICE_NAME=rakazo-worker # Prompt, response and tool content stays excluded unless exactly true. OTEL_CAPTURE_CONTENT=false # Opt-in Pi JSONL session recording under DATA_DIR/pi-sessions. Default off: # transcripts are unencrypted on disk and unsuitable for multi-tenant hosted DATA_DIR. # PI_SESSION_RECORDING=true # Turns the ticket board on for every account; otherwise each person turns it on in # Settings > Labs. Set it on both API and worker. TICKET_BOARD_ENABLED=false # Signed-in desktop apps let bots run commands on the user's own computer; false turns it off on API and worker. # USER_DEVICES_ENABLED=false # Optional account-menu links that open a same-origin page in the web app (docs/self-host-clients.md#menu-links). # RAKAZO_NAV_LINKS='[{"id":"clients","label":"Clients","path":"/ext/clients/"}]' # Optional comma-separated emails of verified accounts that review plugin submissions, add stdio # MCP servers, and grant free access at /admin/access. # RAKAZO_ADMIN_EMAILS=admin@example.com # Optional bot calls and hold-to-talk; enable on both API and worker. PHONE_CALLS_ENABLED=false # Minimum minutes between bot calls to the same user (default 2). PHONE_CALL_COOLDOWN_MINUTES=2 # Seconds before an unanswered call expires (default 45). PHONE_CALL_RING_SECONDS=45 # Per-file chat/computer upload limit in MiB (positive integer; default 10). MAX_UPLOAD_SIZE_MB=10 # Chat attachments per message (integer 1–100; default 4). MAX_ATTACHMENTS_PER_MESSAGE=4 # Off by default. Replies longer than this are read aloud as a summary from the bot's model. # VOICE_SUMMARY_MIN_CHARS=600 WAKEUP_DRIVER=graphile # Keep app-owned PostgreSQL pools bounded. Leave DB_POOL_MAX unset so each # process keeps its default (api 4, worker 8). A set value applies to both — # size it for the worker (three LISTEN clients + leadership + query headroom; # Graphile requires at least 6). Agent runs share this pool between operations. # DB_POOL_MAX=8 # Concurrent legacy interactive queue deliveries during rolling upgrades. # New agent runs dispatch directly from Postgres and do not occupy Graphile slots. # Background queue dispatch and maintenance each have one additional worker. GRAPHILE_WORKER_CONCURRENCY=4 # Model stream retries: integer 0-5; default 1. Invalid or blank values use 1. # More retries allow longer turns during transient provider failures. MODEL_STREAM_MAX_RETRIES=1 # Model timeouts in whole milliseconds: request 1000-2147483647, stream idle 0-2147483647, # connection 1000-300000. Set stream idle to 0 to disable inactivity detection. # Invalid or blank values use the default. Raise them when a shared local model server queues requests. # Wait for a response to start (default 120000; OpenAI Chat Completions/Responses/Codex, Anthropic # Messages, Mistral and Azure OpenAI APIs; others such as Google, Vertex AI, Bedrock and some # routed models use their client's own wait), # for the next streamed chunk (default 180000, every provider), and to connect to an # OpenAI-compatible model or speech connection added in the app (default 30000; RAKAZO_LOCAL_MODELS # servers keep Node's own connect wait). # MODEL_STREAM_TIMEOUT_MS=120000 # MODEL_CONNECT_TIMEOUT_MS=30000 # MODEL_STREAM_IDLE_TIMEOUT_MS=180000 # Characters a message_user progress update may carry before it is cut off: positive # integer; default 500. Mid-turn narration keeps the 500 cap. # RAKAZO_MESSAGE_USER_MAX_CHARS=500 # Bot-to-bot messages one chain may carry before the next is refused: positive integer; # default 6. A person's own message starts a fresh chain. # BOT_MESSAGE_MAX_HOPS=6 # How long a queued run may sit before the worker sends a "still waiting" # reminder (default 14400000, 4 hours) and before it stops the run and frees the bot's computer (default # 86400000, 24 hours). The expiry also stops a question whose run holds a login or code of its own; # other questions stay open. Milliseconds, 60000 to 2592000000; anything else keeps the default. # STUCK_WORK_NOTIFY_AFTER_MS=14400000 # STUCK_WORK_EXPIRE_AFTER_MS=86400000 # Pause or stop computers after this many idle ms. Minimum 30000; 0 disables idle sleep. # Set 0 when self-hosting with Docker, where an idle computer costs nothing, to keep long-running apps and sessions up. # E2B, CreateOS and Boat still apply their own timeouts. SANDBOX_IDLE_MS=600000 # Maximum uptime on providers that bill while running. Minimum 60000; 0 disables the cap. COMPUTER_MAX_ACTIVE_MS=7200000 # Maximum uptime on providers that do not bill while running (Docker, Kubernetes, Incus). Minimum 60000; 0 (default) disables the cap. # Like COMPUTER_MAX_ACTIVE_MS, it needs idle sleep on: SANDBOX_IDLE_MS=0 also turns it off. COMPUTER_MAX_ACTIVE_UNBILLED_MS=0 # Comma-separated model ids on centrally managed local / openai-compatible endpoints that accept images. # User-connected OpenAI-compatible models use the per-connection Supports images setting instead. # Declared ids get input ["text","image"] so screenshot computer tools stay available. RAKAZO_LOCAL_VISION_MODELS= RAKAZO_OPENAI_COMPATIBLE_VISION_MODELS= # Comma-separated local model ids that accept OpenAI-style reasoning_effort. # Declared ids get thinking levels; others never receive the field. RAKAZO_LOCAL_REASONING_MODELS= SANDBOX_COMMAND_TIMEOUT_MS=300000 # team | dedicated: the computer a new bot gets when its request names none. Default team; # dedicated gives each new bot a Private Computer. Existing bots keep theirs. # COMPUTER_DEFAULT_MODE=team # Optional limit on active bot desktops per Docker Team Computer. # 0 leaves capacity to the machine; desktops and Chrome start only when needed. SANDBOX_TEAM_SCREEN_LIMIT=0 # Optional cap on Computer records still backing live bots, per user. Each record # becomes a sandbox container once provisioned, so the cap bounds the containers # one user can create. 0 means no configured cap (default). SANDBOX_MAX_COMPUTERS_PER_USER=0 # Optional limit on total concurrent bot computer containers per space in Docker sandboxes. # 0 (or unset) means unlimited. Set a positive integer to cap containers per space. SANDBOX_MAX_COMPUTERS_PER_SPACE=0 # Docker sandboxes only. "restricted" keeps computer egress to the public internet only # (blocks the host, LAN/RFC1918, and cloud metadata endpoints) and needs # infra/compose/restrict-computer-egress.sh run once on the Docker host; computers refuse # to start until it has. Compose defaults to auto: restricted when the host rules are in # place, otherwise open. A host-run supervisor (pnpm dev, usually Docker Desktop, which # cannot take the rules) defaults to open. See docs/self-host-docker-computers.md. # SANDBOX_COMPUTER_EGRESS=open # Docker sandboxes only. Comma-separated DNS server IPs for bot computers, instead of the # host's resolvers. With restricted egress, set public resolvers here when Docker is older than # 28 and the host's resolver is private, or when the daemon's `dns` option names a private one: # the ruleset blocks those lookups. Under restricted or auto egress, when the ruleset is in place, # computers start only once one of these servers answers. See docs/self-host-docker-computers.md. # SANDBOX_COMPUTER_DNS=1.1.1.1,1.0.0.1 # Docker sandboxes only. Host directories mounted into every bot computer in every Space, as # comma-separated source:target[:rw][:rslave] entries, read-only unless rw. Targets sit below /mnt, # /media or /srv. For a network share that can be remounted, mount it below a stable source # directory and add rw:rslave. See docs/self-host-docker-computers.md. # SANDBOX_COMPUTER_MOUNTS=/srv/shared:/mnt/shared # Docker sandboxes only. Host GPUs every bot computer in every Space may compute on: all or a # count, as in `docker run --gpus`. The Docker host needs a GPU driver for containers (for # example a vendor's container toolkit). See docs/self-host-docker-computers.md. # SANDBOX_COMPUTER_COMPUTE_GPUS=all # Docker sandboxes only. Proxy auto-config (PAC) script for the bot browser, so selected sites # go through an HTTP proxy and the rest connect directly: an absolute http(s):// URL, or the # script as a base64 data: URL. Unset changes nothing. See docs/self-host-docker-computers.md. # SANDBOX_COMPUTER_BROWSER_PROXY_PAC_URL=https://proxy.example.com/proxy.pac # Optional per-turn tool-call fuse for the Pi agent runtime. Unset, empty, or 0 # means unlimited (default). Set a positive integer to soft-stop a turn that # exceeds the budget and still emit a final assistant message. MAX_TOOL_CALLS_PER_TURN= # Deployment-wide fallback model. PI_DEFAULT_PROVIDER picks which key below is used. # A provider that authenticates from the host (amazon-bedrock with an ECS task, IRSA or EC2 # instance role, and AWS_REGION) needs no key: set PI_DEFAULT_CREDENTIALS=host, plus # PI_DEFAULT_PROVIDER and PI_DEFAULT_MODEL. Runs then use and are billed to those credentials. # PI_DEFAULT_CREDENTIALS=host OPENROUTER_API_KEY= ANTHROPIC_API_KEY= DEEPSEEK_API_KEY= PI_DEFAULT_PROVIDER=openrouter # Vision-capable model for chat and computer screenshots. Leave unset to use the # chosen provider's default model (see docs/self-host-models.md). # Computer-use tests can override this with COMPUTER_E2E_MODEL. # PI_DEFAULT_MODEL=anthropic/claude-haiku-5.5 # Optional Auto Review verifier. Default is the deployment or user LLM (no extra key). # Users choose their own approval connection, model and backups in Settings → Models. # Optional deployment default through the existing OpenRouter connection: # RAKAZO_AUTO_REVIEW_PROVIDER=openrouter # RAKAZO_AUTO_REVIEW_MODEL=typesafe/jev-1.14 # Or use TypeSafe directly instead of the LLM checker. # RAKAZO_AUTO_REVIEW_PROVIDER=jev # TYPESAFE_API_KEY= # Optional Jev model (default jev-latest) and minimum choice confidence (0-1, default 0.5). # Below the confidence floor, Auto Review asks the user. # RAKAZO_AUTO_REVIEW_MODEL=jev-latest # RAKAZO_AUTO_REVIEW_MIN_CONFIDENCE=0.5 # Optional Fish Audio speech model used when a connection does not set its own. # Default is s2.1-pro. A Fish connection can override this in Voice settings. # FISH_TTS_MODEL=s1 # Optional local OpenAI-compatible models. Leave blank to disable the provider. RAKAZO_LOCAL_MODELS= RAKAZO_LOCAL_MODELS_URL=http://127.0.0.1:11434/v1 # Optional key for a local server that requires one (LiteLLM, vLLM --api-key, a gateway). # Sent as the bearer token; unset sends the placeholder "local". RAKAZO_LOCAL_MODELS_API_KEY= RAKAZO_LOCAL_CONTEXT_WINDOW=32768 RAKAZO_LOCAL_MAX_TOKENS=4096 # Optional: prompt tokens per minute the local server accepts. Calls are paced under it instead of # waiting out a 429. Unset sends every call at once. # RAKAZO_LOCAL_TOKENS_PER_MINUTE=200000 # Set to off when the local server does not cache prompts. Stale page state is then trimmed from # the prompt as soon as it is worth it, instead of being kept to protect a cache that is not there. # RAKAZO_LOCAL_PROMPT_CACHE=off # Set to 0 to block user-connected OpenAI-compatible endpoints on public hostnames. RAKAZO_OPENAI_COMPAT_ALLOW_PUBLIC= # Comma-separated exact model hostnames allowed to resolve only to private addresses. # No URLs or wildcards; metadata and link-local addresses remain blocked. RAKAZO_OPENAI_COMPAT_PRIVATE_HOSTS= # Set to 1 to let bot credentials (request_secret) target plain-HTTP origins on # private LAN hosts (10.x, 172.16-31.x, 192.168.x, 100.64-127.x, *.local, loopback). RAKAZO_SECRETS_ALLOW_PRIVATE_HTTP= # Allow remote MCP, API/GraphQL connectors, and OpenAI-compatible model and speech servers # on loopback / RFC1918 / Docker-network hosts for every user. # No app account bypasses this policy. Leave unset on public installs. # Self-hosted voice is hidden from settings when this is unset and # RAKAZO_OPENAI_COMPAT_ALLOW_PUBLIC=0. Its connection details are entered in the UI. MCP_ALLOW_PRIVATE_ENDPOINT= # Stdio MCP servers (off unless MCP_STDIO_ENABLED=true and the command is allowlisted). MCP_STDIO_ENABLED= # Comma-separated exact commands, for example: npx,/opt/mcp/bin/server MCP_STDIO_ALLOWED_COMMANDS= # `npx ` servers are installed once here, shared by the API and worker. # Default: DATA_DIR/mcp-packages (or /data/mcp-packages), else the OS temp dir. MCP_PACKAGES_DIR= # Install timeout (default 180000), handshake timeout after launch (default 30000), and how # long an unused server process keeps running (default 900000; 0 keeps it until shutdown). MCP_STDIO_INSTALL_TIMEOUT_MS= MCP_STDIO_STARTUP_TIMEOUT_MS= MCP_STDIO_IDLE_MS= # How long a run waits for one MCP server's tool list before going on without it (default 20000). MCP_DISCOVERY_TIMEOUT_MS= MCP_MAX_TOOLS=250 MCP_MAX_PAGES=20 E2B_API_KEY= COMPOSIO_API_KEY= # Optional integrations.sh-compatible catalog base URL (Rakazo appends api/search). # Example: https://integrations.sh/ or a self-hosted mirror base, not the full search endpoint. INTEGRATIONS_CATALOG_URL= # Optional alternative managed connector provider. PIPEDREAM_CLIENT_ID= PIPEDREAM_CLIENT_SECRET= PIPEDREAM_PROJECT_ID= PIPEDREAM_ENVIRONMENT=development # Optional messaging surface (Chat SDK). Each platform mounts when its full # credential set is present; webhooks arrive at /api/v1/messaging/webhook/. # Users link a chat app to their account from the web (Messaging settings). # Set true to instead let unknown senders auto-provision their own chat-only # accounts (their runs bill the deployment model key): MESSAGING_OPEN_SIGNUP=false # Sendblue (iMessage/SMS — all four required): SENDBLUE_API_KEY_ID= SENDBLUE_API_SECRET= SENDBLUE_SIGNING_SECRET= SENDBLUE_PHONE_NUMBER= # Slack (both required). Channels and rooms use the shared messaging team-chat # path when TEAM_CHAT_BOT_ID is set; Slack is one adapter, not a separate stack. # The bot token needs the reactions:write scope for the 👀 receipt. # Multi-workspace installs: configure the Slack app OAuth redirect as # /api/v1/messaging/install/slack/callback. No bot token is needed. SLACK_CLIENT_ID= SLACK_CLIENT_SECRET= # Single-workspace fallback: SLACK_BOT_TOKEN= SLACK_SIGNING_SECRET= # Gateway bot for provider-neutral team/external rooms (Slack, later Teams, etc.). # Alias SLACK_RAKAZO_BOT_ID is accepted for older configs. TEAM_CHAT_BOT_ID= # Optional separate judge model for ambient room replies (both or neither): TEAM_CHAT_JUDGE_PROVIDER= TEAM_CHAT_JUDGE_MODEL= # Microsoft Teams direct messages (single-tenant bot). # Messaging endpoint: https://your-domain.example/api/v1/messaging/webhook/teams TEAMS_APP_ID= TEAMS_APP_PASSWORD= TEAMS_APP_TENANT_ID= # WhatsApp Business Cloud (all four required): WHATSAPP_ACCESS_TOKEN= WHATSAPP_PHONE_NUMBER_ID= WHATSAPP_APP_SECRET= WHATSAPP_VERIFY_TOKEN= # Telegram (both required). Inbound is webhook-only at # /api/v1/messaging/webhook/telegram with header X-Telegram-Bot-Api-Secret-Token; # the worker must not long-poll getUpdates. TELEGRAM_BOT_TOKEN= TELEGRAM_WEBHOOK_SECRET_TOKEN= # Extra bots: TELEGRAM_BOT_TOKEN_ + TELEGRAM_WEBHOOK_SECRET_TOKEN_ # (letters/digits). Provider: telegram-; webhook: /api/v1/messaging/webhook/telegram-. # Feishu/Lark (app id, secret, and verification token required). Inbound is # webhook-only at /api/v1/messaging/webhook/lark — do not enable long # connection in the developer console or the HTTP webhook will miss events. # Create a custom app with the Bot capability, subscribe to im.message.receive_v1, # and set the event subscription URL to that path. LARK_DOMAIN=feishu uses # open.feishu.cn (default); lark uses open.larksuite.com. Encrypt key is # optional unless event encryption is enabled in the console. LARK_APP_ID= LARK_APP_SECRET= LARK_VERIFICATION_TOKEN= LARK_ENCRYPT_KEY= LARK_DOMAIN= # iMessage through Photon (https://photon.codes), project id and secret required. # Inbound arrives over Spectrum's outbound stream, so no public URL is needed. # With PHOTON_WEBHOOK_SECRET set, Spectrum posts signed webhooks to # /api/v1/messaging/webhook/photon instead and nothing streams. PHOTON_PROJECT_ID= PHOTON_PROJECT_SECRET= PHOTON_WEBHOOK_SECRET= # Extra projects: PHOTON_PROJECT_ID_ + PHOTON_PROJECT_SECRET_ # (+ optional PHOTON_WEBHOOK_SECRET_), letters/digits. Each project gives # a person its own number, so each can reach a different bot. Provider: photon-. # Mobile push. Set only when the Expo project enables enhanced push security. EXPO_ACCESS_TOKEN= # Optional transactional email: resend (HTTP API), smtp, or none. # Unset EMAIL_PROVIDER preserves SMTP_URL-based configuration. EMAIL_PROVIDER= EMAIL_API_KEY= SMTP_URL= EMAIL_FROM= EMAIL_REPLY_TO= # Optional bot email addresses (name@INBOUND_EMAIL_DOMAIN). A mail relay posts each message to # /api/v1/email/inbound with this secret; see docs/self-host-inbound-email.md. Set both or neither. INBOUND_EMAIL_DOMAIN= INBOUND_EMAIL_SECRET= # Separate TLS SMTP connection for bot mail. Unset disables sending, independently of account mail. BOT_EMAIL_SMTP_URL= # Optional product mail: a welcome email and, when referrals are enabled, a day-one referral email. WELCOME_EMAIL_ENABLED= # Local development only: capture account email without contacting a provider. EMAIL_EMULATOR= # Billing (optional; off when unset). Set both secrets together to enable model credits. # BILLING_EMULATOR=true is an offline stand-in for local development; real keys take priority. # BILLING_EMULATOR= # Use the same Stripe account and mode for the API key and webhook endpoint. # Webhook: /api/v1/billing/webhook; snapshot events, API 2026-09-30.endive. # Events: customer.subscription.*; customer.deleted; checkout.session.completed; # checkout.session.async_payment_succeeded; invoice.paid; invoice.payment_failed; # charge.refunded; charge.dispute.created; charge.dispute.closed. # STRIPE_SECRET_KEY=sk_test_replace-me # STRIPE_WEBHOOK_SECRET=whsec_replace-me # Optional: display the legacy seat price. Credit tiers create reusable prices automatically. # STRIPE_PRICE_ID=price_replace-me # Optional affiliate program: reports signups from partner links and their own-key plan payments. # DUB_API_KEY=dub_replace-me # Structured logs to stdout (json in production, pretty in development unless overridden). LOG_LEVEL=info LOG_FORMAT= # Optional Axiom ingest. Needs both token and dataset. Failures never affect the product. AXIOM_TOKEN= AXIOM_DATASET= # Optional region: AXIOM_EDGE is a hostname (eu-central-1.aws.edge.axiom.co). # AXIOM_EDGE_URL is an https URL and wins when both are set. AXIOM_EDGE= AXIOM_EDGE_URL= # Days to keep security events (sign-ins, credential changes, refusals). Default 90. SECURITY_EVENT_RETENTION_DAYS= # Days to keep the audit log (who did what in each organization). Default 365. AUDIT_EVENT_RETENTION_DAYS= # Optional: delete chat messages older than this many days from chats with nothing running. # Unset keeps them forever. See docs/self-host-configuration.md. CHAT_HISTORY_RETENTION_DAYS= # Optional: forward audit events to an HTTPS endpoint (a SIEM collector) as JSON Lines. AUDIT_LOG_WEBHOOK_URL= # Optional Authorization header value for that endpoint, e.g. "Bearer ". AUDIT_LOG_WEBHOOK_AUTHORIZATION= # Error reporting is off unless a DSN is set. SENTRY_DSN= SENTRY_BROWSER_DSN= SENTRY_ENVIRONMENT= SENTRY_RELEASE= # Optional. GET /internal/health on the API port returns this as `revision` so you can confirm the deployed commit. GIT_SHA= # --- Compose production deployments --- # Absolute path of this checkout as the Docker *daemon* sees it. Production Compose defaults to # /srv/rakazo on Linux; override it for any other host path so the updater bind mount resolves. # Linux override: RAKAZO_DEPLOY_DIR=/opt/rakazo # Docker Desktop on Windows (C:\Users\you\rakazo): # RAKAZO_DEPLOY_DIR=/run/desktop/mnt/host/c/Users/you/rakazo RAKAZO_DEPLOY_DIR= # Compose files the updater reconciles, in overlay order, separated like Compose's own # COMPOSE_FILE (":" unless COMPOSE_PATH_SEPARATOR says otherwise). Leave unset for the base stack. # RAKAZO_COMPOSE_FILE=infra/compose/docker-compose.prod.yml:ops/compose/overlay.yml # Docker sandboxes on the prod stack (instead of a remote provider) layer # infra/compose/docker-compose.prod.docker.yml — list it here too: # RAKAZO_COMPOSE_FILE=infra/compose/docker-compose.prod.yml:infra/compose/docker-compose.prod.docker.yml RAKAZO_COMPOSE_FILE= # Extra services an update pulls, recreates and rolls back, comma separated. Name any overlay # service built from the application image, or it keeps running the previous code after an update. # Appended to api, worker, web; the updater itself is refused. # RAKAZO_UPDATE_SERVICES=supervisor RAKAZO_UPDATE_SERVICES= # Published images (defaults match this repository). Tag `local` is built from the checkout. RAKAZO_IMAGE=ghcr.io/elie222/rakazo/app RAKAZO_IMAGE_TAG=local RAKAZO_IMAGE_TAG_PREVIOUS= RAKAZO_UPDATER_IMAGE=ghcr.io/elie222/rakazo/updater RAKAZO_UPDATER_IMAGE_TAG=local # Memory caps for the containers that run agent turns (Compose sizes, default 2g each). Raise them # on a larger host; a deploy briefly runs the old and new container of each side by side. RAKAZO_API_MEMORY_LIMIT= RAKAZO_WORKER_MEMORY_LIMIT= # How many API and worker containers the production stack runs (default 1 each). Each one uses # about one CPU core; deploys apply a new count. RAKAZO_API_REPLICAS= RAKAZO_WORKER_REPLICAS= # Production Compose Postgres memory, sized for an 8 GB host (defaults 2560m, 1GB, 2GB); see # docs/self-host-configuration.md#postgres-memory for smaller hosts. POSTGRES_MEMORY_LIMIT= POSTGRES_SHARED_BUFFERS= POSTGRES_EFFECTIVE_CACHE_SIZE= # Required only when starting the opt-in `updater` Compose profile; the sidecar refuses to start # without it. Use a dedicated random value, not BETTER_AUTH_SECRET, # SANDBOX_SUPERVISOR_TOKEN, or SCREEN_PROXY_SECRET; # ≥32 chars in production. Leave empty when the sidecar is disabled. RAKAZO_UPDATER_TOKEN= # Optional default UI locale for the web SPA (for example en, de, tr, or pl). # Overridden by localStorage key rakazo.uiLocale when the user picks a language in Settings. # VITE_DEFAULT_UI_LOCALE=en # Expo: English, Simplified Chinese, Russian, or German (en | zh-CN | ru | de). # Overridden by SecureStore key rakazo.uiLocale. Account → Language offers English # and Simplified Chinese; the other catalogs follow this default or the device language. # EXPO_PUBLIC_DEFAULT_UI_LOCALE=en # Optional bot media generation. Models must be explicitly enabled. MEDIA_GENERATION_PROVIDER=disabled MEDIA_GENERATION_BASE_URL=https://queue.fal.run MEDIA_GENERATION_API_KEY= MEDIA_GENERATION_IMAGE_MODEL= MEDIA_GENERATION_VIDEO_MODEL= # Optional paid data services for bots (treg | fake | disabled). Calls run on this one # deployment account; with model credits on, organizations pay cost plus 20% from credits. PAID_SERVICES_PROVIDER=disabled PAID_SERVICES_API_KEY= # PAID_SERVICES_BASE_URL=https://treg.to # Optional content-free product analytics (API and worker). Off unless selected. # A/B tests to run, comma-separated keys from packages/contracts/src/experiments.ts. None by default. # EXPERIMENTS= # none | posthog | emulator (local development only: events go to the API log). ANALYTICS_PROVIDER=none ANALYTICS_KEY= # Generate once (openssl rand -hex 32); share between API and worker and retain on token rotation. # Error reports also use it to count affected users by a hashed id. ANALYTICS_PSEUDONYM_SECRET= ANALYTICS_HOST=https://us.i.posthog.com