External APIs and data

Free chat uses the same server-side Chat Completions API, sending the current text and the player's own history trimmed according to that player's settings; it reuses the existing account, billing and limits. In this mode no tools or web search are called this round. See free chat.

This page is compiled from the current repository code; checked 2026-10-30. It describes the external requests made by this mod's built-in implementation and does not cover the network behavior of Minecraft, the loader or other mods, nor models, TTS, ASR or payment implementations registered by third-party extensions themselves.

Overview

Feature	Request initiator	External service	When it is requested	Default state
AI intent detection, sign-in reward selection, special quest text organization, quest background filtering	Minecraft server; the internal server in single player	The configured OpenAI Chat Completions-compatible service, DeepSeek by default	When the model is actually needed and the current player allows external AI	The default provider is DeepSeek; the server must configure a key
Web search and fetch	Minecraft server worker thread	The separately configured DeepSeek Search / Fetch API	When the local database query is empty, the search module is enabled and the current request explicitly allows network access	Not granted automatically; an ordinary question does not count as authorization; reuses the DeepSeek environment variable
Text-to-speech (TTS)	Minecraft client	Alibaba Cloud Bailian / DashScope	Narrating the current message when auto-read is on, or a manual preview	Auto-read is off by default
Speech-to-text (ASR)	Minecraft client	Alibaba Cloud Intelligent Speech Interaction NLS	When the player turns on voice input and records actively	Off by default
NLS temporary token issuance	Minecraft client	Alibaba Cloud NLS Meta CreateToken	When an active recognition has no reusable token, the token is about to expire or the authorization cache is invalid	Triggered on demand with voice input; not requested when settings are saved
Non-streaming TTS audio download	Minecraft client	The Alibaba Cloud OSS address named in the TTS response	When non-streaming synthesis is called and an audio URL is returned	Normal in-game narration uses streaming PCM and does not need this download
Voice design, cloning, query, deletion and temporary upload	Standalone Python tool	Alibaba Cloud Bailian and OSS	When the corresponding subcommand of tools/voice_design.py is run manually	Not executed automatically while the mod runs

FTB Quests queries, submissions and reward claims call the local server's Java API; there is no separate HTTP or stdio MCP service. Lightweight intent matching also runs locally and does not request an embedding API.

FTB guide search uses the local sparse TF-IDF on the server; the guide button on the quest page and /system guide do not request the chat or embedding API. Ordinary Q&A goes into system:question, and authorized passages found locally can be handed to the text model for an answer; the index, the whole corpus, other players' data and progress are not sent as a whole. Vague natural language may still go through model intent detection first, and when auto-narration is on the existing TTS is still called.

1. Server text model API

Address and protocol

• Method: POST; requests and responses are JSON, currently with stream=false.
• Path: /chat/completions is appended to baseUrl from config/system/system-model.json.
• Default full address: https://api.deepseek.com/chat/completions, default model deepseek-chat.
• For example, with https://example.com/v1 as baseUrl the actual request is https://example.com/v1/chat/completions; do not append /chat/completions again in the config.
• It can be replaced with another OpenAI-compatible service. Remote addresses must use HTTPS, while an HTTP service on a loopback address is also allowed.
• Authentication: the server environment variable named by apiKeyEnvironment is read, SYSTEM_DEEPSEEK_API_KEY by default, and placed in Authorization: Bearer .... When a local service without a key is allowed, apiKeyRequired can be turned off.

DeepSeek's thinking and reasoning_effort compatibility only changes the parameters of the same request and does not add another API request. For the model, JSON mode and output token parameters see Model API.

Which behaviors call the model

Call scenario	Main content sent to the model	Cases where the request can be avoided
Intent routing fallback	The current input, a limited amount of recent conversation verbatim and a summary of earlier results, local candidate modules with similarity, descriptions of available modules and their input schema, the player language, the routing prompt	Structured actions such as buttons, explicit commands, high-confidence local read-only matches or a valid intent cache hit
Free chat / extension fallback	The current message, a limited amount of the player's own chat history, the player language, the base chat prompt and the persona TXT body referenced by the currently bound System; full free mode additionally includes the available tool descriptions and this turn's observations	Offline / standard mode, the administrator disabling free chat or the model, the player turning AI off; a successful route in extension mode does not enter the chat fallback
AI sign-in reward selection	The current player's name, health and maximum health, experience level, dimension, language, date and the candidate reward definitions from the data pack	When the player has already signed in that day, the recorded result is returned directly and nothing is selected again
Narration organization of special quest requirements or rewards	Objective or reward types that cannot be formatted directly, the native title, objective counts and known limits, plus the language and the organizing prompt	Common objectives are formatted directly; an organized passage cache hit; the corresponding narration is not enabled
Local knowledge Q&A organization	The current question, the player language, matched passages from the current player's private / shared knowledge and authorized FTB guides, their titles and categories	Nothing is organized when the database query is empty, external AI is off or no model budget remains; on a recoverable failure the sources are shown without searching
Web retrieval decision and organization	The current question, the player language, this request's source summaries and a limited amount of page text, the remaining tool budget	It runs only after the local result is empty, the search module is enabled and this request explicitly allows network access
Quest background filtering	The quest title, the raw visible FTB description, the language and the background filtering prompt	Detail or background narration is off; the description is empty, the title is duplicated or the description is only a URL; a valid background cache hit

All built-in model requests go through ExecutionContext.model → ModelGateway → OpenAiCompatibleClient and share permissions, quotas, concurrency limits and usage statistics. One natural-language operation may first call routing and then the model logic of the feature itself; not every message results in exactly one request.

The sign-in model receives only a dedicated projection of the current player and is not sent the online player list or the full inventory. Routing may send the player's recent input verbatim and restricted fields from earlier results such as quest, reward or claim references; text the player enters on purpose may also contain personal information. The whole world save, the entire chat history or all player records are never packaged and sent automatically.

Special quest organization and background filtering use only the quest definitions and do not send the player's inventory, current quest progress or a whole conversation; the quest text itself may still contain content written by the data pack author. Background results are stored in the system:quest_backgrounds table of system/database/shared.dat in the current world save (GZIP-compressed NBT) and are reused by quest ID, language and content fingerprint; a successful "no background" conclusion is cached as well.

After the player selects the "offline" mode, these model calls and server-side web search are all disabled; local queries and existing caches can still be used. Voice features have their own switches and keys, and the offline chat mode does not turn off client TTS or ASR together.

Implementation entry points:

• Model HTTP client
• Model connection config
• Intent routing
• Sign-in model input
• Quest narration organization / Background filter cache

2. Client TTS API

The config file is the client-side config/system/system-speech.json. The key comes from apiKey in the config; when it is empty the client environment variable DASHSCOPE_API_KEY is read. The client requests Alibaba Cloud directly and does not have the Minecraft server send the request on its behalf.

Domain selection

Region	No Workspace ID	Workspace ID set
Beijing beijing	https://dashscope.aliyuncs.com	https://<workspace>.cn-beijing.maas.aliyuncs.com
Singapore singapore	https://dashscope-intl.aliyuncs.com	https://<workspace>.ap-southeast-1.maas.aliyuncs.com

Synthesis API

Model	Method and path
The default qwen3-tts-flash and the date versions the adapter allows	POST /api/v1/services/aigc/multimodal-generation/generation
qwen-audio-3.1-tts-flash, qwen-audio-3.1-tts-plus, qwen-audio-3.0-tts-flash, qwen-audio-3.0-tts-plus	POST /api/v1/services/audio/tts/SpeechSynthesizer

Authentication uses Authorization: Bearer .... The request sends the actual text to be read, the model, the voice ID, the synthesis language or language hint; Qwen-Audio also sends the format and sample rate. The text may contain a host name, quest titles, quest requirements, rewards or an allowed background, depending on the current message and the voice settings; a whole chat history is never passed to TTS automatically. Using an existing custom voice sends only the voice ID, and normal synthesis does not upload the cloning sample again.

Current in-game narration sends X-DashScope-SSE: enable and receives the SSE / PCM audio stream within the same HTTP request, instead of starting a new synthesis request for each audio fragment. The final audio URL returned by the API is not used for streaming playback. Auto-read is off by default, and a manual preview calls the synthesis API actively; the local notification sound does not call TTS.

The optional whole-sentence audio cache is off by default; once enabled and hit, the local audio is played directly without calling the synthesis API and without adding cloud usage. Muting, a missing key or another state that does not allow playback skips normal automatic synthesis. Closing the screen or confirming a notification cancels the current request and playback, but a request that has already reached the provider cannot be guaranteed to produce no usage.

Retained non-streaming audio download

The non-streaming synthesize uses the same synthesis API. If the response gives an audio URL, the client sends a separate GET to download the WAV; if Base64 audio is returned directly no download is needed.

The download address must be an Alibaba Cloud OSS domain the adapter allows (the host ends with .aliyuncs.com and contains .oss-), and an HTTP address is upgraded to HTTPS. The download carries no Bailian Bearer key and does not follow redirects. This path exists in the code; current normal in-game narration prefers the SSE streaming path described above.

Implementation entry points: Alibaba Cloud TTS adapter, synthesis request fields, narration session.

3. Client ASR API

• Protocol: WebSocket over TLS, address wss://nls-gateway-cn-<region>.aliyuncs.com/ws/v1.
• Currently supported regions: shanghai, beijing, shenzhen, Shanghai by default; for example wss://nls-gateway-cn-shanghai.aliyuncs.com/ws/v1.
• Config: client config/system/system-asr.json, using a separate NLS project appKey and Alibaba Cloud accessKeyId / accessKeySecret, not the Bailian TTS API key.
• When the config is empty, ALIYUN_NLS_APP_KEY, ALIYUN_AK_ID and ALIYUN_AK_SECRET are read respectively. The automatically issued token goes into the handshake request header X-NLS-Token, and the AppKey goes into the recognition command header.
• Issuance: GET https://nls-meta.cn-shanghai.aliyuncs.com/ with Action CreateToken, Version 2019-02-28 and RegionId fixed to cn-shanghai. It sends the AccessKey ID, a UTC timestamp, a unique nonce and the format and signature parameters; the AccessKey Secret is only used locally to compute HMAC-SHA1 and is not put into the request directly.
• The token is cached in memory only; ExpireTime is checked at actual recognition time and a new token is issued when two minutes or less remain; a credential change triggers a new fetch. On a 401 / 403 handshake the cache is cleared, a new token is issued and one reconnect happens, before the recognition command has been submitted; a billed task that has already started is not retried. Issuance failures and handshake failures do not count as ASR attempts.
• The same connection sends SpeechRecognizer.StartRecognition, PCM audio frames and StopRecognition, and receives intermediate results and the final recognized text.

Voice input is off by default; once enabled, the recording session opens only when the player actively starts voice input. It sends this microphone recording (16 kHz, 16-bit, mono PCM) plus the AppKey, task ID, message ID and parameters needed for recognition, and does not send the whole chat history. A single recording is currently at most 60 seconds, and closing the related screen or notification cancels the session.

After a successful recognition the text fills the input box, or is submitted directly according to the "send immediately after recognition" setting; notification speech first matches the current notification buttons and only falls back to the main chat handling when nothing matches. If AI intent detection is needed afterwards, the server model requests of section 1 may still occur; ASR itself is not a model routing request.

Token issuance happens only in a recognition session the player started actively, and opening the settings or saving credentials does not access the network; the signed URL, Secret, token and upstream body are never written to the log. Simple Voice Chat is only used as an optional recording device source; without it Java Sound is used, and the ASR audio is uploaded to NLS directly by this mod's client. Only when the player sends the recognized text or performs a notification action does it enter the normal assistant request; the AccessKey, token and ASR / TTS usage are not uploaded to the Minecraft server.

Implementation entry points: Token issuance and caching, Alibaba Cloud ASR adapter, ASR config, Recording and recognition session.

4. APIs used by the standalone voice tool

The following requests happen only when tools/voice_design.py is run manually; they do not run automatically with MC startup, opening the settings page or ordinary message narration. The tool can read the client voice config, but it is not an in-game background task.

Operation	Method and path / target	Content sent
design voice design	POST /api/v1/services/audio/tts/customization, model voice-enrollment, action create_voice	Text description of the voice, preview text, target model, voice prefix, language and audio parameters
clone voice cloning	The same customization API, action create_voice	Base64 data of the selected WAV, target model, prefix, language, sample duration and preprocessing options
list / query / delete	The same customization API, with actions list_voice / query_voice / delete_voice	The prefix and page number, or the specified voice ID
synthesize test synthesis	POST /api/v1/services/audio/tts/SpeechSynthesizer	Test text, model, voice and WAV parameters; when a URL is returned, an additional GET to OSS downloads it
upload get a temporary upload policy	GET /api/v1/uploads?action=getPolicy&model=<model>	The target model and the Bailian Bearer key
upload upload a file	POST to the HTTPS OSS upload_host returned in the previous step	The selected audio file and the signed form fields, without the Bailian Bearer key

Voice management and synthesis reuse the domain selection of section 2; the tool's voice management operations are currently restricted to the Beijing region. Temporary upload policies use the compatibility domains dashscope.aliyuncs.com or dashscope-intl.aliyuncs.com and do not switch to the workspace domain even when a Workspace ID is given.

clone submits the sample as Base64 directly and does not call the temporary upload first. upload is a separate subcommand that returns an oss:// resource address; uploading does not mean a voice has been created or synthesis has been completed. The tool's key comes from the specified config, DASHSCOPE_API_KEY or a hidden input, and is not written into the exported voice config.

5. Server DeepSeek web search and fetch API

The feature module system:web_search exports only the Search/Fetch tools and calls DeepSeekSearchClient through SearchGateway. The core system:question always queries the local database first; only when the result is empty, the search module is enabled and this request is explicitly authorized does it enter WebResearch with a budget. /search <question>, an explicit request for network access or a submit from the search page allows the fallback, while /question and ordinary questions do not. For natural language the model still decides the current authorization, and history and material text do not grant permission; a database query error, a hit whose organization fails, sign-in, binding and events never start a search. The search module is not added to players automatically.

• Config: config/system/system-search.json, copied from the JAR defaults on first use; it is independent of the text model config and still reads the environment variable SYSTEM_DEEPSEEK_API_KEY by default.
• Search: POST https://api.deepseek.com/v1/tools/search, sending the query derived from the current question, a fixed 5 results and options without body text; built-in workflows do not fill the raw question and location fields and do not collect player data or the IP automatically.
• Fetch: POST https://api.deepseek.com/v1/tools/fetch, sending the page URL provided by the caller to DeepSeek, which fetches the Markdown; the local machine does not access that page directly.
• Search card links are handed to Minecraft's vanilla link flow only after the player clicks them, and the system browser is opened according to the link switch and confirmation settings; the browser then visits the target site on its own, and the mod does not attach the DeepSeek key to that URL. Expanding a source summary sends no network request.
• Authentication: Authorization: Bearer ... is sent only to the configured API address, the key is not attached to result URLs or icon addresses, and results are not fetched automatically.
• Search / Fetch are each estimated at 5 CNY per thousand successful requests that returned results and go into the player's world usage ledger separately; an empty result is not priced, and unknown results are listed separately. Model tokens are counted separately. All messages use Codec, and the network reuses the HttpTransport that limits length, applies timeouts, forbids redirects and does not retry automatically.

See web search and fetch for detailed configuration, response fields, calling methods and official material.

6. Which operations do not access external APIs additionally

• Player and item queries, FTB queries / explicit submission / reward claims, starter gift claiming, shop and lottery settlement: the built-in logic reads or modifies the server game state directly; executing these operations through an explicit command or a trusted button does not itself need the cloud. When natural language first goes through the model fallback, or the result triggers TTS or special quest narration, the corresponding requests may still occur.
• Binding success notifications, quest completion / unlock events and the message queue: event generation and fixed text are handled locally; with voice on, TTS and on-demand quest text organization may be triggered separately.
• Semantic matching and conversation compression: use local lexical / TF-IDF, caches and deterministic history projections; there is no embedding, vector database, Redis, separate summarization model or automatic model download request.
• Usage statistics and cost estimates: AI usage is recorded by the server from model responses, and Search/Fetch counts and estimates are recorded by the server gateway; TTS / ASR usage is stored on the local machine by client account only, and rates are read from the config under the client config/system; provider billing, balance or real-time pricing APIs are never requested, and voice usage is not reported to the MC server.
• Prompt and reward data pack reload, config loading and notification sound playback: these read local resources or the world save and do not call the cloud.

Assistant requests, card actions and AI usage queries between the client and server use Minecraft custom network payloads and are not a new public HTTP API. The TTS key, the NLS AccessKey / token and voice usage all stay on the client and are not sent to the game server.

According to the current built-in source, this mod has no additional telemetry, advertising, update check or general data upload interface. The external request implementations are concentrated in OpenAiCompatibleClient, AliyunTtsProvider, AliyunNlsToken, AliyunAsrProvider, DeepSeekSearchClient and the Python tool above; after adding a provider or another third-party extension this list should be updated as well.

Scope of the check

This list is compiled from the request calls, address concatenation, configs and serialized fields in the source. No provider API was actually requested, no key was used for connectivity or billed calls, and the game was not started nor tests run. Domains and paths reflect the current code; the availability, quotas and billing of provider APIs are subject to the actual service.