Skip to content

Configuration Reference

Every setting hypeman-social reads, with defaults. Nothing here is required to import the library — configuration is read lazily, when a component authenticates or generates.

How lookup works

Plain configuration — get_config(section, key)

Priority, first match wins:

  1. Doppler — if DOPPLER_TOKEN is set, the sectioned key (SETTINGS_CHECK_INTERVAL) then the simple key (CHECK_INTERVAL) are looked up in the Doppler project/config. The project is fetched once per process and shared with the secret lookups below; reset_secret_cache() forces a re-read.
  2. Simple env key — CHECK_INTERVAL
  3. Sectioned env key — SETTINGS_CHECK_INTERVAL
  4. The default passed by the caller.

Empty strings count as unset — a blank line in .env never overrides a default. A value that Doppler returns and that matches the placeholder rule is skipped.

Booleans accept true, 1, yes, on, enabled (case-insensitive); anything else is false. Unparseable ints/floats fall back to the default with a logged warning.

This document lists keys in their sectioned form (LLM_PROVIDER, BLUESKY_HANDLE, …), which is the convention all four daemons use. The simple form (PROVIDER) also resolves, but don't — it collides across sections.

Secrets — get_secret(platform, key)

Credentials go through a stricter chain, so a production secrets manager genuinely overrides a stale local .env:

  1. Doppler (if DOPPLER_TOKEN is set)
  2. AWS Secrets Manager (if SECRETS_MANAGER=aws)
  3. HashiCorp Vault (if SECRETS_MANAGER=vault)
  4. Environment / .env — <PLATFORM>_<KEY>, e.g. BLUESKY_APP_PASSWORD
  5. The default.

Required secrets and the placeholder rule

The library never starts a feature on a missing or template credential. One helper, hypeman_social.config.is_placeholder(value), defines the rule for every lookup (get_secret, platform_secret, get_env_secret, constructor credential overrides, and secrets-manager bundles). A value counts as unset when, case-insensitively and ignoring surrounding space, it is:

  • empty, or whitespace only;
  • wrapped in angle brackets, such as <token>;
  • exactly example or placeholder;
  • starting with your_, your-, changeme, change_me, change-me, xxx or dp.st.your;
  • a URL (or other /, ?, =, :, @ separated value) with any segment matching the above, so https://discord.com/api/webhooks/1/YOUR_WEBHOOK_TOKEN is caught.

A placeholder behaves exactly like an unset value: the feature stays disabled, no request goes to the real service, and the log names the variable (for example DISCORD_WEBHOOK_URL), never its value. Real values are returned unchanged.

Secret Needed for Without it Where to put it
DISCORD_WEBHOOK_URL or DISCORD_WEBHOOK_<PLATFORM> Discord posting Discord posting stays disabled env, Doppler, AWS, Vault
DISCORD_BOT_TOKEN Discord chat receiving Chat adapter unavailable same
MATRIX_ACCESS_TOKEN, or MATRIX_USERNAME + MATRIX_PASSWORD Matrix posting Matrix stays disabled same
BLUESKY_APP_PASSWORD Bluesky posting Bluesky stays disabled same
MASTODON_ACCESS_TOKEN Mastodon posting Mastodon stays disabled same
THREADS_ACCESS_TOKEN Threads posting Threads stays disabled same
TWITCH_CLIENT_ID, _ACCESS_TOKEN, _REFRESH_TOKEN, _CLIENT_SECRET Twitch chat Chat unavailable, error names the key same
LLM_CLAUDE_API_KEY or ANTHROPIC_API_KEY Claude provider Provider disabled same
LLM_OPENAI_API_KEY or OPENAI_API_KEY OpenAI provider Provider disabled same
LLM_GEMINI_API_KEY or GEMINI_API_KEY Gemini provider Provider disabled same

Only the features you enable need their secrets. The daemons built on this library inherit the rule; leave a variable empty rather than copying a sample value into it.

load_config(env_path=".env") loads a .env file into the environment; call it once at startup (repeat calls are cheap no-ops).

Secrets managers

All optional. With none configured, everything reads from env / .env.

Variable Meaning Default
DOPPLER_TOKEN Doppler service token; presence enables Doppler —
DOPPLER_PROJECT Doppler project —
DOPPLER_CONFIG Doppler config prd
SECRETS_MANAGER aws, vault, or none none
SECRETS_VAULT_URL Vault server URL —
SECRETS_VAULT_TOKEN Vault token —

Per-platform bundle locations (all optional; a conventional <platform> bundle name is tried when unset):

Variable pattern Example
SECRETS_AWS_<PLATFORM>_SECRET_NAME SECRETS_AWS_BLUESKY_SECRET_NAME=prod/bluesky
SECRETS_VAULT_<PLATFORM>_SECRET_PATH SECRETS_VAULT_BLUESKY_SECRET_PATH=secret/bluesky
SECRETS_DOPPLER_<PLATFORM>_SECRET_NAME SECRETS_DOPPLER_BLUESKY_SECRET_NAME=BLUESKY

AWS credentials/region come from the standard boto3 chain (AWS_ACCESS_KEY_ID, AWS_DEFAULT_REGION, instance profiles, …).

Install the matching extra: hypeman-social[aws], [vault], [doppler].

LLM: core

Variable Meaning Default
LLM_ENABLE Master switch. Off = LLMManager.authenticate() returns False and callers use their template fallbacks false
LLM_MAX_TOKENS Response token budget per generation 150
LLM_TEMPERATURE Sampling temperature. Low by default because small models get "creative" and invent giveaways 0.3
LLM_TOP_P Nucleus sampling 0.9
LLM_MAX_RETRIES Attempts per generation before giving up 3
LLM_RETRY_DELAY_BASE Seconds before first retry; doubles each attempt 2

Thinking mode (reasoning models: qwen3, gemma4, deepseek-r1)

Variable Meaning Default
LLM_ENABLE_THINKING_MODE Give reasoning models headroom to think, and mine the thinking text when the answer field comes back empty false
LLM_THINKING_TOKEN_MULTIPLIER Token budget multiplier while thinking mode is on 4.0

When the connected Ollama model advertises the thinking capability, hypeman passes an explicit think parameter on every generation: think=false forces a direct answer, think=true routes reasoning into the response's thinking field where extraction can reach it. Models (and old servers) without the capability never see the parameter. With thinking mode off (the default) a thinking-capable model is sent think=false, so a one-sentence chat answer is not spent on hidden reasoning.

LLM: provider selection and failover

Variable Meaning Default
LLM_PROVIDER Primary provider: ollama, gemini, claude (alias anthropic) or openai gemini
LLM_FALLBACK_PROVIDER Ordered, comma-separated failover chain, e.g. gemini or gemini,ollama. Strictly opt-in: if you chose Ollama so your data stays local, hypeman will never silently ship prompts to Google (none)
LLM_MODEL Model name shared by all providers provider default
LLM_OLLAMA_MODEL / LLM_GEMINI_MODEL / LLM_CLAUDE_MODEL / LLM_OPENAI_MODEL Per-provider model; wins over LLM_MODEL. Needed whenever primary and fallback run different model names provider default

Claude and OpenAI read only their own namespace. Every setting of those two providers (LLM_CLAUDE_*, LLM_OPENAI_*, plus the shared LLM_MODEL) is looked up by that exact name, Doppler first and then the environment. Unlike the other providers, the bare names (OPENAI_BASE_URL, OPENAI_MODEL, CLAUDE_MODEL, CLAUDE_EFFORT, ...) are ignored, so a variable set for other tooling cannot redirect the endpoint that receives your API key. claude and anthropic name the same provider, so listing both in LLM_FALLBACK_PROVIDER builds one client. A malformed number (LLM_*_TIMEOUT, LLM_*_RATE_LIMIT) logs a warning naming the setting and uses the default; it never stops startup. That includes the shared numeric settings these two providers inherit (LLM_MAX_TOKENS, LLM_MAX_RETRIES, LLM_TEMPERATURE, ...); the other providers still fail fast on a typo.

Where the key goes. What is guaranteed, and what is not:

  • Both clients are built with an explicit base URL (https://api.anthropic.com, https://api.openai.com/v1, or your validated LLM_OPENAI_BASE_URL), so ANTHROPIC_BASE_URL / OPENAI_BASE_URL in the environment are never used.
  • Redirects are disabled: a 3xx is an error, never followed with the key.
  • The HTTP client ignores the process environment for network trust (trust_env=False): HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, NO_PROXY, SSL_CERT_FILE, SSL_CERT_DIR and REQUESTS_CA_BUNDLE have no effect. Anyone who controls those could otherwise terminate TLS for the pinned host with their own CA and read the key. For a private proxy or CA use the settings below; both are read by exact name (Doppler, then env) and an invalid value makes that provider unavailable with one error naming the setting, never a silent fallback to the environment.
Variable Meaning
LLM_CLAUDE_PROXY / LLM_OPENAI_PROXY http:// or https:// proxy URL with a host. Credentials in the URL are refused (they would land in config and logs), as are query and fragment
LLM_CLAUDE_CA_BUNDLE / LLM_OPENAI_CA_BUNDLE Path to a readable PEM file. It replaces the default trust store for that provider's client
  • Not covered: someone who can set LLM_* settings, edit the config source (Doppler project, .env), or run code in the process controls the key anyway. The headers OPENAI_ORG_ID / OPENAI_PROJECT_ID and OPENAI_CUSTOM_HEADERS / ANTHROPIC_CUSTOM_HEADERS are still read by the SDKs and sent to the pinned host only (custom headers can replace an auth header, so do not set them in this process's environment unless you mean to). ANTHROPIC_AUTH_TOKEN and Anthropic profile files are not consulted because the API key is passed explicitly.
  • Error handling: provider errors carry only a fixed safe reason (class, status, "connection error", ...), never SDK text, and are raised after the SDK handler exits so they have no chained exception on the provider's own path. If your code calls the provider from inside an except block, Python attaches that outer exception; the provider marks it suppressed so formatted tracebacks do not print it, but the object stays reachable. Object level secrecy is not claimed. The API key also stays on the provider instance (reconnect needs it), so a traceback's frame locals or a debugger can reach it; do not ship full-locals tracebacks (for example sentry with include_local_variables) from this process.
  • Structured-output enums are limited to 256 plain-JSON options, 8 levels, 2000 nodes and 16,384 characters of strings and object keys in total; a larger enum is refused before the call.

The manager keeps re-probing a downed primary in the background and switches back automatically when it recovers — you stop paying for the cloud fallback the moment your local box is back.

LLM: Ollama

Install extra: hypeman-social[ollama].

Variable Meaning Default
LLM_OLLAMA_HOST Server host; http:// is assumed when no scheme given http://localhost
LLM_OLLAMA_PORT Port, appended only when the host has none 11434
LLM_OLLAMA_MODEL Model, e.g. gemma4:12b, gemma3:4b, qwen3:8b gemma4:12b
LLM_OLLAMA_NUM_CTX Context size, sent as options.num_ctx exactly as written. Leave unset to send nothing. On a shared GPU, a num_ctx different from the one the model is loaded with makes Ollama reload it and can evict another resident model, so set this only to the value your server already uses (not sent)
LLM_OLLAMA_KEEP_ALIVE Passed as keep_alive when set (e.g. 30s, 5m), so a non-resident model can be given a short residency (not sent)

A model missing from the server logs a warning but doesn't fail — Ollama pulls it on first use.

LLM: Gemini

Install extra: hypeman-social[gemini].

Variable Meaning Default
LLM_GEMINI_API_KEY API key (secret chain applies; GEMINI_API_KEY also accepted) —
LLM_GEMINI_MODEL Model name gemini-2.0-flash-lite
LLM_GEMINI_RATE_LIMIT Client-side requests/minute self-throttle, so you never discover the free-tier limit via 429s 30

LLM: Claude

Install extra: hypeman-social[claude] (the Anthropic SDK needs Python 3.10 or newer; on 3.9 the extra installs nothing and the provider reports itself unavailable). Set LLM_PROVIDER=claude (or put claude in LLM_FALLBACK_PROVIDER). Built for high-volume chat, so the defaults are the cheap path.

Variable Meaning Default
LLM_CLAUDE_API_KEY API key (secret chain applies, Doppler first; ANTHROPIC_API_KEY also accepted) —
LLM_CLAUDE_MODEL claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5 or claude-fable-5-1 claude-haiku-5-5
LLM_CLAUDE_SYSTEM_PROMPT Stable persona/rules prefix sent as the system block. Keep it byte-identical between calls (none)
LLM_CLAUDE_PROMPT_CACHE Put a cache_control breakpoint on the system block true
LLM_CLAUDE_CACHE_TTL 5m or 1h 5m
LLM_CLAUDE_TIMEOUT Per-request timeout, seconds 30
LLM_CLAUDE_EFFORT_LEVEL low, medium, high, xhigh, max, or omit for the model default low
LLM_CLAUDE_THINKING auto, disabled, adaptive, between_tools, omit. auto sends disabled on Haiku and omits it elsewhere (Sonnet and Opus 5.5 reject disabled) auto
LLM_CLAUDE_RATE_LIMIT Client-side requests/minute self-throttle; 0 disables 50

What the options cost:

  • Prompt caching is the main saving when LLM_CLAUDE_SYSTEM_PROMPT is long and calls are frequent: cache reads bill at about 0.1x the input rate, the first write at 1.25x (5 minute TTL) or 2x (1 hour TTL). A prefix below the model's minimum cacheable length silently does not cache, so a short persona gains nothing; check usage.cache_read_input_tokens. Use 1h only when calls are minutes apart.
  • Effort and thinking spend output tokens. low effort with thinking off keeps a chat reply near its visible length. Raising either costs output tokens, billed at the output rate, and thinking shares the max_tokens cap.
  • Sampling parameters (temperature, top_p) are never sent, because Haiku 5.5 and Sonnet 5.5 reject non-default values. LLM_TEMPERATURE and LLM_TOP_P therefore do not apply to Claude; steer from the prompt.
  • Limits and retries: the SDK's own retries are off, so LLM_MAX_RETRIES and LLM_RETRY_DELAY_BASE are the only retry layer. A safety decline (stop_reason: refusal) returns no text, which the manager treats like any other failed generation (next provider, or the template).

Haiku 5.5 is $0.10 / $0.50 per million input / output tokens up to 100K prompt tokens; check the Anthropic pricing page before committing a budget.

LLM: OpenAI

Install extra: hypeman-social[openai] (the OpenAI SDK needs Python 3.10 or newer; on 3.9 the extra installs nothing and the provider reports itself unavailable). Uses the Responses API.

Variable Meaning Default
LLM_OPENAI_API_KEY API key (secret chain applies, Doppler first; OPENAI_API_KEY also accepted) —
LLM_OPENAI_MODEL Model id gpt-6-luna
LLM_OPENAI_SYSTEM_PROMPT Sent as instructions (none)
LLM_OPENAI_REASONING_EFFORT none, low, medium, high, xhigh, max, or omit for the model default low
LLM_OPENAI_TIMEOUT Per-request timeout, seconds 30
LLM_OPENAI_BASE_URL Only for an OpenAI-compatible gateway; unset uses OpenAI. Must be https (plain http only for localhost/loopback) with no user info, query or fragment; anything else fails authentication and the key is never attached to it (none)
LLM_OPENAI_RATE_LIMIT Client-side requests/minute self-throttle; 0 disables 60

Requests are sent with store: false, and temperature/top_p are not sent (current models do not document them). Reasoning tokens are billed as output tokens, so low keeps chat cheap; raise it only for work that needs it. gpt-6-luna is listed at $0.10 / $0.50 per million input / output tokens; check the OpenAI pricing page before committing a budget.

Structured output. OpenAILLM.generate_structured(prompt, schema, name) sends a strict JSON schema and returns the parsed object, or None. None covers a model refusal, an incomplete (truncated) response, text that is not JSON (duplicate keys, NaN and Infinity count as invalid), and any mismatch with your schema, checked locally: type, properties, required, additionalProperties, enum (at most 256 plain-JSON options, 8 levels deep) and items; integer accepts 1.0 and 1e0 as JSON Schema does. A schema using any other keyword (pattern, minimum, oneOf, ...) is refused before the call, because the result could not be verified. It never repairs or guesses, so a safety classifier should treat None as "no verdict" and fail closed. It shares availability, throttling, retry and error handling with generate(). Strict mode requires additionalProperties: false and every property listed in required. Schemas add a small first-call latency.

LLM: guardrails

Applied by apply_guardrails() after generation; a failed hard check returns None and the daemon posts its template instead.

Variable Meaning Default
LLM_ENABLE_DEDUPLICATION Reject messages too similar to recent ones (>80% word overlap). Lives on the manager, so failover doesn't wipe the history true
LLM_DEDUP_CACHE_SIZE Recent messages remembered 20
LLM_MAX_EMOJI_COUNT More than this reads as spam 2
LLM_ENABLE_PROFANITY_FILTER Reject profanity false
LLM_PROFANITY_SEVERITY mild, moderate, or severe — which word lists to check. moderate checks mild+moderate words; severe checks everything moderate
LLM_ENABLE_QUALITY_SCORING Score 0–10 against generic-phrase and relevance heuristics false
LLM_MIN_QUALITY_SCORE Minimum passing score 6
LLM_ENABLE_PLATFORM_VALIDATION Per-network checks (@everyone on Discord, inline URLs on Bluesky, HTML entities on Mastodon) true

Not configurable — always on: hallucination detection from the active content profile (invented star counts, "premiering tonight", "drops enabled", …), forbidden hype-word list, exact hashtag-count enforcement, and stripping hashtags derived from your own username.

LLM: recovery

The part that exists because of the outage.

Variable Meaning Default
LLM_ENABLE_AUTO_RECONNECT Heal a downed provider on the next availability check true
LLM_RECONNECT_INTERVAL Cooldown between reconnect attempts (seconds) 60
LLM_MAX_RECONNECT_ATTEMPTS Give up after N attempts; 0 = never give up 0

Social: Bluesky

Install extra: hypeman-social[bluesky].

Variable Meaning Default
BLUESKY_ENABLE_POSTING Enable the platform false
BLUESKY_HANDLE Account handle, e.g. you.bsky.social —
BLUESKY_APP_PASSWORD App password (secret chain applies) — create one at Settings → App Passwords, never your account password —

Behavior: 300-grapheme limit enforced with real grapheme counting (ZWJ emoji count as one), URLs become link facets, #hashtags become tag facets, link cards are built from stream metadata or an Open Graph scrape, replies thread under the root post. stream_data['images'] attaches up to four pictures with alt text as an images embed (which replaces the link card — Bluesky allows one embed per post).

Social: Mastodon

Install extra: hypeman-social[mastodon].

Variable Meaning Default
MASTODON_ENABLE_POSTING Enable the platform false
MASTODON_API_BASE_URL Your instance, e.g. https://infosec.exchange —
MASTODON_ACCESS_TOKEN Access token (secret chain) — Preferences → Development → New Application, scopes read, write:statuses, write:media —
MASTODON_CLIENT_ID OAuth client id (secret chain). Optional; only useful together with the secret —
MASTODON_CLIENT_SECRET OAuth client secret (secret chain). Optional —

Behavior: replies thread; thumbnails and images upload as media attachments with alt text; a failed upload never blocks the post. The access token alone is enough to post — the client id/secret pair is passed through when both are set and ignored (with a warning) when only one is.

Social: Discord

No extra needed — plain webhooks over HTTP.

Variable Meaning Default
DISCORD_ENABLE_POSTING Enable the platform false
DISCORD_WEBHOOK_URL Default webhook (secret chain) —
DISCORD_WEBHOOK_<PLATFORM> Per-source webhook: DISCORD_WEBHOOK_TWITCH, _YOUTUBE, _KICK, _TIKTOK route each source to its own channel falls back to default
DISCORD_ROLE Role id to mention on posts —
DISCORD_ROLE_<PLATFORM> Per-source role mention falls back to default
DISCORD_ENDED_MESSAGE Text shown when a live embed flips to "stream ended" built-in
DISCORD_ENDED_MESSAGE_<PLATFORM> Per-source ended text falls back to default

Behavior: rich embeds colored per source (Twitch purple, YouTube red, Kick green, TikTok cyan), live embeds are edited in place with fresh viewer counts and thumbnails, and flip to a muted "ended" embed keeping the VOD link. The event kind (upload vs live) decides embed style — pass it explicitly (see API).

Chat: Discord (receiving)

Needs the discord-chat extra (websocket-client). Setup, the privileged intent and the bridge notes are in DISCORD_CHAT.md. Every key goes through the usual secret chain (Doppler first, then the environment), as DISCORD_<KEY>.

Variable Meaning Default
DISCORD_BOT_TOKEN Bot token (secret chain). Never logged required
DISCORD_GUILD_ID Only hear this guild all guilds the bot is in
DISCORD_CHANNEL_IDS Comma-separated channel ids to hear all text channels the bot can see
DISCORD_MAX_EVENT_BYTES Largest inbound Gateway frame or message, in bytes. An event over it stops the receiver with one error naming this setting. Positive integer, at most 67108864; anything else fails closed 2097152 (2 MiB)
DISCORD_MODERATOR_ROLE_IDS Comma-separated role ids that map to the moderator role none (nobody is a moderator)

Social: Matrix

No extra needed — client-server HTTP API via requests.

Variable Meaning Default
MATRIX_ENABLE_POSTING Enable the platform false
MATRIX_HOMESERVER Homeserver URL; https:// assumed —
MATRIX_ROOM_ID Target room, e.g. !abc123:matrix.org —
MATRIX_USERNAME + MATRIX_PASSWORD Bot login — preferred: a fresh access token per start, tokens rotate automatically —
MATRIX_ACCESS_TOKEN Static token — used only when username/password are absent —

Behavior: HTML-formatted messages with clickable links and per-source headers; replies use m.in_reply_to. Matrix cannot edit messages, so no live viewer-count updates.

Social: Threads

No extra needed — the Threads Graph API is plain HTTPS.

Variable Meaning Default
THREADS_ENABLE_POSTING Enable the platform false
THREADS_ACCESS_TOKEN Long-lived user access token —
THREADS_USER_ID Numeric Threads user ID —

Behavior: two-step publish (create a media container, then publish it); 500-character limit enforced with URL-preserving trims; the first URL in the message is attached as an explicit link-preview card; replies thread via reply_to_id.

Getting credentials: create a Meta app at developers.facebook.com, add the Threads API use case, grant threads_basic and threads_content_publish, then generate a long-lived access token for your account. Your numeric user ID comes back from GET https://graph.threads.net/v1.0/me?fields=id. Long-lived tokens last ~60 days; refresh them with GET .../refresh_access_token before expiry.

Chat: Twitch

Needs the twitch extra (pip install "hypeman-social[twitch]", which adds websocket-client). Hears one channel and replies as a dedicated bot account. Setup, scopes and revocation are in TWITCH_BOT.md.

Every key is read through the secret chain (Doppler first) as platform_secret('Twitch', key); there are no plain settings.

Key Meaning Required
client_id Client ID of the bot's own Twitch application yes
client_secret That application's client secret; used only to refresh the token for refresh
access_token The bot account's user access token (user:read:chat, user:write:chat) yes
refresh_token The matching refresh token for refresh
bot_user_id Numeric Twitch user id of the bot account yes
channel Channel to join, as a login name or a numeric user id yes

The adapter shares the TWITCH_ prefix with any other Twitch credentials in the same Doppler config (for example a stream announcer's own application). Point the bot at its own bundle with SECRETS_DOPPLER_TWITCH_SECRET_NAME (for example TWITCHBOT, so keys read TWITCHBOT_CLIENT_ID), or give it its own Doppler config.

Logging

Optional; use configure_logging() from hypeman_social.observability.

Variable Meaning Default
LOG_LEVEL DEBUG / INFO / WARNING / ERROR INFO
LOG_FILE Path; enables rotating file output (stdout only)
LOG_MAX_BYTES Rotate at this size 10485760 (10 MB)
LOG_BACKUP_COUNT Rotated files kept 5
LOG_TO_STDOUT Also log to stdout true
LOG_TIMESTAMPS Include timestamps (auto-off under systemd, where journald stamps already) auto
LOG_DEDUPE_SECONDS Collapse identical repeats within N seconds 0 (off)
LOG_QUIET_LIBRARIES Quiet chatty HTTP libraries true

Worked examples

Local-first AI with cloud failover, posting everywhere

LLM_ENABLE=true
LLM_PROVIDER=ollama
LLM_OLLAMA_HOST=http://192.168.1.50
LLM_OLLAMA_MODEL=gemma4:12b
LLM_FALLBACK_PROVIDER=gemini
LLM_GEMINI_MODEL=gemini-2.0-flash-lite
GEMINI_API_KEY=...

BLUESKY_ENABLE_POSTING=true
BLUESKY_HANDLE=you.bsky.social
BLUESKY_APP_PASSWORD=...

MASTODON_ENABLE_POSTING=true
MASTODON_API_BASE_URL=https://your.instance
MASTODON_CLIENT_ID=...
MASTODON_CLIENT_SECRET=...
MASTODON_ACCESS_TOKEN=...

DISCORD_ENABLE_POSTING=true
DISCORD_WEBHOOK_URL=https://discord.com/api/webhooks/...

MATRIX_ENABLE_POSTING=true
MATRIX_HOMESERVER=https://matrix.org
MATRIX_ROOM_ID=!room:matrix.org
MATRIX_USERNAME=@bot:matrix.org
MATRIX_PASSWORD=...

Fully local, nothing leaves the network

LLM_ENABLE=true
LLM_PROVIDER=ollama
LLM_OLLAMA_HOST=http://localhost
LLM_OLLAMA_MODEL=gemma4:12b
# no LLM_FALLBACK_PROVIDER: an Ollama outage means template posts, not Google

Production: Doppler holds the secrets

DOPPLER_TOKEN=dp.st....
DOPPLER_PROJECT=stream-daemon
DOPPLER_CONFIG=prd
# every key above can live in Doppler instead of this file,
# and Doppler wins over anything set locally