Services are the functional units that power every step of an agent workflow. Each service represents a distinct capability — from answering questions and extracting data, to querying databases, running classifiers, calling external APIs, and generating reports.

Only services that have been added and enabled in the Services section of an agent are available for selection inside nodes. This keeps agents modular and predictable: you can see exactly what each agent is capable of at a glance, and adding or removing a service directly controls what nodes can do.

Services are organized into five categories based on their purpose. The table in each section serves as a quick reference — the descriptions beneath provide the context you need to use each one effectively.

1. Core services

General building blocks for common tasks such as chat, retrieval, extraction, and logic control.

1.1 chat

The general-purpose model interaction node. Accepts a question or instruction and generates a response from the active language model. Works off MAIN-INPUT, User-Text, or no context at all — making it suitable for drafting, summarizing, reformatting, or any open-ended task at any point in the pipeline.

Instruction: What is your question or instruction?
Context: MAIN-INPUT, User-Text, None

1.2 rag_batch

Answers a question across a collection of documents rather than a single file. It searches the entire User-Collection batch, retrieves the most relevant passages from each, and constructs a grounded response. Best suited for research pipelines and multi-file document review workflows.

Instruction: Enter question or instruction
Context: User-Document

1.3 rag_answer

Designed for long single-document inputs where the source exceeds standard prompt limits. rag_answer chunks the document internally, retrieves the most relevant sections, and returns an answer grounded in the actual text — not the model's general knowledge.

Instruction: Ask question to longer document input
Context: User-Source, Provide_instruction_or_query

1.4 vision

Adds visual reasoning to the workflow. Supply an image, ask a question, and the service responds based on what the multimodal model sees — reading embedded text, identifying objects, describing scenes, or interpreting diagrams.

Instruction: Enter question to image file
Context: User-Image

1.5 ocr_vision

For documents and images where both text extraction and visual context matter. Applies OCR to extract the text first, then uses vision-based reasoning to interpret it in context — suited for scanned forms, annotated diagrams, and mixed-content PDFs.

Instruction: Enter instruction
Context: User-Document, User-Image

1.6 ocr

Extracts text from image-based or copy-protected documents and registers the result as a named source in the agent state. The extracted content then becomes available to any downstream node — for RAG, extraction, or summarization — just like any regular document.

Instruction: Enter name of new document source
Context: User-Document

1.7 agent_report

Reads the accumulated agent state and compiles everything into a structured, titled report. Typically placed at the end of a workflow. No specific input context is required — it assembles from whatever the pipeline has produced up to that point.

Instruction: Enter title for agent report
Context: -

1.8 wikipedia_search

Fetches Wikipedia article content and injects it into the agent state as research context. No pre-loaded documents are needed, making it a lightweight way to add factual grounding to any workflow without managing external sources.

Instruction: Add Wikipedia Articles as Research Context
Context: None

1.9 prompt_builder

Constructs a structured prompt dynamically from the current agent state rather than relying on a hardcoded string. Best placed immediately before a model-execution node to ensure the prompt is well-formed and adapts to the pipeline's state at runtime.

Instruction: Enter prompt instruction
Context: None

1.10 embedded_bot

Pauses workflow execution and surfaces an interactive interface for a human operator. Execution resumes once the operator provides input or confirms a decision. The instruction field is optional. This is the primary mechanism for human-in-the-loop pipelines.

Instruction: Optional
Context: None

1.11 condition

Evaluates a logical expression against the current agent state and routes execution accordingly. Acts as an if/else branch point in the pipeline — no external context required. The expression can reference any key present in the agent state.

Instruction: Enter expression
Context: None

1.12 web_search

Queries the web at runtime and returns structured results — titles, snippets, and source URLs — for downstream nodes to act on. Useful when the workflow requires up-to-date or real-time information that is not available in any pre-loaded source.

Instruction: Add query
Context: None

1.13 boolean

Takes a yes/no question, evaluates it against the input text, and returns a binary answer along with a brief explanation. Best placed at screening or validation steps where a clear true/false signal is needed before the workflow proceeds.

Instruction: Provide yes/no question
Context: MAIN-INPUT, User-Text

1.14 extract

Pulls a specific named value from unstructured text. Specify the key as the instruction, and the service returns the corresponding value from the input. Use it to surface discrete fields — dates, names, amounts, identifiers — and carry them forward in the agent state.

Instruction: Enter extraction key
Context: MAIN-INPUT, User-Text

1.15 answer

A context-constrained alternative to chat. Responds to a factual question using only the supplied input text — never the model's broader training knowledge. Use it wherever responses must be directly traceable to the provided passage.

Instruction: What is your question?
Context: MAIN-INPUT, User-Text

2. Classifiers

Lightweight text analysis tools for labeling or scoring content.

2.1 sentiment

Runs sentiment analysis on the input text and returns a positive, negative, or neutral classification. No instruction is needed — the service operates directly on MAIN-INPUT or User-Text. Useful for customer feedback pipelines, review analysis, and any workflow where emotional tone needs to be assessed at scale.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.2 emotions

Goes a level deeper than sentiment by identifying the specific emotion present in the text — such as joy, frustration, surprise, or fear. Like all classifiers, it requires no instruction and runs directly on MAIN-INPUT or User-Text.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.3 topics

Identifies the primary topic or subject area of the input text. No instruction is required. Useful for categorizing incoming content, routing documents to the right downstream nodes, or tagging records in bulk processing workflows.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.4 tags

Generates a set of descriptive tags from the input text. No instruction is required. Tags can be used to index content, enable filtering in later nodes, or surface key themes across a batch of documents.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.5 intent

Classifies the intent behind the input text — for example, whether the user is making a request, asking a question, expressing a complaint, or providing information. No instruction is needed. Particularly useful in customer-facing workflows or message triage pipelines.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.6 ratings

Scores the positivity of the input text on a 1–5 scale. No instruction is required. Provides a numeric signal that can be carried forward in the agent state, used in conditions, or aggregated across a batch for reporting purposes.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.7 ner

Performs Named Entity Recognition on the input text, identifying and labeling entities such as people, organizations, locations, dates, and other proper nouns. No instruction is required. Outputs structured entity data that can be extracted and used in downstream nodes.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.8 xsum

Generates an extreme summary — a single, highly compressed sentence that captures the core meaning of the input. No instruction is required. Use it when you need the most concise possible distillation of a passage, such as for indexing, previews, or high-volume batch summarization.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.9 summary

Produces a standard prose summary of the input text. The instruction field is optional — leaving it empty generates a general summary, while providing a specific instruction can focus the summary on a particular aspect of the content.

Instruction: Optional
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.10 category

Assigns the input text to a predefined category. No instruction is required. The category label is returned as a structured output in the agent state, where it can be used in condition nodes, filters, or downstream reporting.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

2.11 q_gen

Generates a set of questions based on the input text. No instruction is required. Useful for creating evaluation datasets, populating FAQs, building comprehension assessments, or surfacing potential knowledge gaps in a document.

Instruction: No instruction required
Context: MAIN-INPUT, User-Text, Transformer Node (with selection of a previous output from an earlier agent step from a chat or vision bot, or dataset column that contains relevant text)

3. Datasets

Tools for preparing, querying, and analyzing structured data.

3.1 build_dataset

Creates a structured dataset from a JSON input. Provide a name for the dataset in the instruction field — the service parses the incoming JSON and registers the result as a named dataset in the agent state, making it available for all subsequent dataset operations.

Instruction: Enter dataset name
Context: JSON Input

3.2 select_keys

Filters a JSON dictionary down to only the keys you specify. Provide the key names in the instruction field, and the service returns a new JSON object containing only those fields. Useful for trimming large JSON objects before passing them into downstream nodes.

Instruction: Enter keys
Context: JSON Input

3.3 dataset_plot

Generates a visual chart or graph from the current dataset. Provide a visualization instruction — such as specifying the chart type or the columns to plot — and the service produces a visual output that can be included in reports or reviewed in the agent interface.

Instruction: Enter visualization instruction
Context: Dataset

3.4 load_dataset

Loads a previously saved dataset back into the agent state by name. Provide the dataset name in the instruction field. Use this when a workflow needs to resume from a saved state or when a dataset created in an earlier run needs to be reused.

Instruction: Enter dataset name
Context: Dataset

3.5 create_json

Consolidates selected keys from the agent state into a single JSON dictionary. Provide the list of keys in the instruction field, and the service assembles them into a structured JSON object. Useful for packaging results before passing them to extraction, export, or integration nodes.

Instruction: Enter keys list
Context: Agent-State

3.6 ds_command_filter

Applies a filter command to the dataset using a structured expression. Provide the filter command in the instruction field. This service is suited for programmatic filtering where precise control over the filter logic is required.

Instruction: Enter filter command
Context: Dataset

3.7 ds_column_filter

Keeps only the rows in the dataset where a selected column meets a specified condition. Provide the column and condition in the instruction field. Useful for narrowing a dataset to a relevant subset before analysis or reporting.

Instruction: Enter column condition
Context: Dataset

3.8 ds_quick_stats

Generates a statistical summary report for a selected column — including metrics like count, mean, min, max, and distribution. Provide the column name in the instruction field. Designed for fast exploratory analysis without running a full statistical pipeline.

Instruction: Select column
Context: Dataset

3.9 ds_column_analysis

Produces a detailed analytical report for a selected column, covering distribution, patterns, and outliers. More thorough than ds_quick_stats. Provide the column name in the instruction field.

Instruction: Select column
Context: Dataset

3.10 ds_report

Generates a combined dataset and workflow report with no additional instruction required. It summarizes the dataset contents alongside the workflow steps that produced them — useful as a closing node for data-heavy pipelines.

Instruction: No instruction
Context: Dataset

3.11 ds_column_select

Returns the values of a single selected column from the dataset. Provide the column name in the instruction field. Use it to isolate a specific column and surface its values for downstream extraction, filtering, or reporting.

Instruction: Select column
Context: Dataset

3.12 ds_ask_dataset

Queries the dataset using a natural language question. Provide the question in the instruction field, and the service interprets the query, retrieves the relevant data, and returns a structured or prose answer. No SQL or filter syntax is required.

Instruction: Enter query
Context: Dataset

3.13 ds_readout

Returns the raw text content from a specified row range in the dataset. Provide the row range in the instruction field. Useful for inspecting specific records, feeding row-level content into downstream model nodes, or sampling data from large datasets.

Instruction: Enter row range
Context: Dataset

3.14 ds_smart_filter

Filters the dataset semantically based on meaning rather than exact keyword or column matching. Provide a natural language query in the instruction field, and the service returns rows whose content is semantically relevant to that query.

Instruction: Enter query
Context: Dataset

3.15 ds_keyword_filter

Filters the dataset by exact keyword match. Provide the keyword in the instruction field, and the service returns all rows containing that term. Best used when precision matters and semantic fuzzy-matching is not appropriate.

Instruction: Enter keyword
Context: Dataset

3.16 ds_statistics

Performs a deeper statistical analysis of the entire dataset — going beyond column-level summaries to examine distributions, correlations, and other aggregate patterns. No instruction is required.

Instruction: No instruction
Context: Dataset

3.17 ds_stat_analysis

Runs a statistical analysis specifically on CSV-format datasets. Similar in scope to ds_statistics but optimized for CSV input structure. No instruction is required.

Instruction: No instruction
Context: Dataset

4. Specialized services

Advanced utilities for targeted or complex workflows.

4.1 build_table

Creates a structured database table from CSV data. Provide a name for the table in the instruction field, and the service registers it in the agent state from a User-Table input. Once built, the table can be queried in natural language by query_custom_table in downstream nodes.

Instruction: Enter table name
Context: User-Table

4.2 query_custom_table

Queries a custom database table using a natural language question. Provide the query in the instruction field, and the service translates it into a structured lookup against the Table Output from a preceding build_table node. No SQL knowledge is required.

Instruction: Enter query
Context: Table Output

4.3 json_extractor

Converts embedded JSON text found within unstructured content into a structured dataset element. Provide a schema in the instruction field to define the expected structure. Works on MAIN-INPUT or User-Text and is useful for processing API responses, log outputs, or any prose that contains inline JSON.

Instruction: Enter schema
Context: MAIN-INPUT, User-Text

4.4 semantic_filter

Filters a source collection based on semantic meaning rather than exact text matching. Provide a natural language instruction describing what content to keep, and the service returns passages from User-Source that are meaningfully relevant to that description.

Instruction: Enter instruction
Context: User-Source

4.5 text_filter

Applies rule-based filtering to a source collection using a keyword or topic. Provide the keyword or topic phrase in the instruction field, and the service returns matching passages from User-Source. Best used when you need deterministic, exact-match filtering rather than semantic relevance.

Instruction: Enter keyword/topic
Context: User-Source

4.6 document_filter

Filters a source collection down to content from a specific named document. Provide the document name in the instruction field, and the service returns only passages originating from that file within the User-Source. Useful when multiple documents are loaded as a source and you need to scope a node's input to just one.

Instruction: Enter document name
Context: User-Source

4.7 table_filter

Applies structured filtering logic to a source collection containing tabular data. No instruction is required — the service applies its own parsing and filtering rules to the User-Source input and returns the relevant structured content.

Instruction: No instruction
Context: User-Source

4.8 load_kb

Loads a pre-built knowledge base into the agent state by name. Provide the knowledge base name in the instruction field, and its content becomes available as context for downstream nodes. No document input is required — the KB is retrieved from storage directly.

Instruction: Enter KB name
Context: None

4.9 ds_ask_kb

Queries a knowledge base using a natural language question derived from a dataset input. Provide the question in the instruction field. The service retrieves relevant answers from the loaded knowledge base and returns them alongside dataset-level context.

Instruction: Enter query
Context: Dataset

4.10 transformer

Applies a text or data transformation to the current agent state. Provide the transformation instruction or choose the input source in the instruction field. Use it to reshape, convert, reformat, or restructure data between nodes without invoking a full model inference step.

Instruction: Choose input
Context: Agent-State

4.11 aggregate_context

Merges multiple named context sources into a single unified context. Provide the names of the contexts to consolidate in the instruction field. Useful when several upstream nodes have produced separate context blocks that need to be combined before being passed to a model or report node.

Instruction: Enter context names
Context: None

4.12 parse_document

Converts a document file into plain text and registers it in the agent state under a given name. Provide the name in the instruction field. Accepts User-Document input and is typically used as a preprocessing step before extraction, RAG, or analysis nodes.

Instruction: Enter name
Context: User-Document

4.13 create_context

Builds a reusable context block from the most relevant passages in a source collection. Provide a guiding instruction in the instruction field to direct which passages are selected, and the service constructs a focused context from User-Source that can be referenced by downstream nodes.

Instruction: Enter instruction
Context: User-Source

4.14 report_commentary

Generates a narrative commentary based on the current agent state. The instruction field is optional — omitting it produces a general-purpose commentary, while providing a specific focus narrows the output. Useful for adding analytical interpretation to a report before it is finalized.

Instruction: Optional
Context: None

4.15 speech_gen

Generates an audio file from a text input. Provide the text to be spoken in the instruction field, and the service returns an audio output. Useful for creating voice narrations, accessibility outputs, or audio summaries as part of a workflow.

Instruction: Enter text
Context: None

4.16 image_gen

Generates an image from a text description. Provide the description in the instruction field, and the service returns a generated image as output. Useful for workflows that produce visual assets alongside written content.

Instruction: Enter description
Context: None

4.17 get_stock_summary

Retrieves a summary of stock information for a given ticker symbol. Provide the ticker in the instruction field, and the service returns relevant market data. Requires no input context and is typically used in financial analysis or market monitoring workflows.

Instruction: Enter ticker
Context: None

4.18 speech

Transcribes a single speech audio file into text. Provide the input reference in the instruction field, and the service returns the transcribed text from the Audio Input source. Use it as a preprocessing step before applying text-based analysis or extraction nodes.

Instruction: Enter input
Context: Audio Input

4.19 speech_batch

Transcribes a collection of speech audio files into text. Provide an instruction to guide the transcription in the instruction field, and the service processes the entire Audio Batch input. Useful for bulk transcription pipelines such as interview analysis or call center review.

Instruction: Enter instruction
Context: Audio Batch

4.20 vision_batch

Answers questions across multiple images by applying visual reasoning to a batch of User-Document inputs. Provide the question or instruction in the instruction field, and the service processes each image in the batch and aggregates the results.

Instruction: Enter instruction
Context: User-Document

4.21 parse_batch

Parses a batch of documents into a unified text source. Provide a guiding instruction in the instruction field, and the service converts each document in the User-Document batch into text, combining them into a single source available to downstream nodes.

Instruction: Enter instruction
Context: User-Document

4.22 extract-tiny

A lightweight alternative to extract for simple key-value extraction tasks. Provide the key in the instruction field, and the service returns the corresponding value from MAIN-INPUT or User-Text. Best used when the extraction is straightforward and minimizing processing overhead matters.

Instruction: Enter key
Context: MAIN-INPUT, User-Text

4.23 website_scraper

Extracts content from a specified web page. Provide the URL in the instruction field, and the service retrieves and returns the page's text content. Operates on allowed websites only and requires no input context — useful for pulling live web content directly into the agent state.

Instruction: Enter URL
Context: None

4.24 extract_table

Extracts structured table data from within a document. Provide the query or description of the table to extract in the instruction field, and the service locates and returns the matching table from the User-Document input as structured data.

Instruction: Enter query
Context: User-Document

5. Integrations

Connect the agent to external systems or hosted models.

5.1 push_to_s3

Uploads data from the agent state to an Amazon S3 bucket. Provide the destination path in the instruction field. No input context is required — the service reads directly from the agent state. Use it to persist workflow outputs to cloud storage as part of an automated pipeline.

Instruction: Enter path
Context: None

5.2 pull_from_s3

Downloads data from an Amazon S3 bucket into the agent state. Provide the source path in the instruction field. No input context is required. Use it to bring external data files or previously stored outputs into a workflow without manual upload.

Instruction: Enter path
Context: None

5.3 connect_library

Establishes a connection to a semantic library by name. Provide the library name in the instruction field. Once connected, the library context becomes accessible to downstream nodes — particularly query_library — enabling semantic search over curated, pre-indexed content collections.

Instruction: Enter library name
Context: None

5.4 query_library

Queries a connected semantic library using a natural language question. Provide the query in the instruction field, and the service retrieves the most relevant content from the Library Context established by a preceding connect_library node.

Instruction: Enter query
Context: Library Context

5.5 get_quote

Retrieves a real-time stock quote for a given symbol. Provide the ticker symbol in the instruction field, and the service returns current price and related market data. No input context is required. Used in financial workflows alongside get_stock_summary or get_company_financials.

Instruction: Enter symbol
Context: None

5.6 get_company_financials

Retrieves financial data for a specified company or ticker. Provide the company name or ticker in the instruction field, and the service returns structured financial information. No input context is required. Suited for investment research and financial reporting workflows.

Instruction: Enter company/ticker
Context: None

5.7 send_email

Sends an email using content assembled from the agent state. Provide the recipient email address in the instruction field and select the context source to use as the email body. Useful for automated notification, report delivery, or alert workflows.

Instruction: Enter email
Context: Select context

5.8 openai_chat

Sends a prompt to the OpenAI chat completion API and returns the response. Provide the instruction in the instruction field, and the service resolves it against the supplied Text Source. Use it when a specific workflow node requires OpenAI's models rather than a locally running model.

Instruction: Enter instruction
Context: Text Source

5.9 openai_rag

Performs a Retrieval-Augmented Generation query using OpenAI's models against a Text Source. Provide the instruction in the instruction field. The service retrieves relevant passages and returns a grounded answer — functionally equivalent to rag_answer but routed through the OpenAI API.

Instruction: Enter instruction
Context: Text Source

5.10 openai_rag_batch

Runs RAG across a batch of text sources using OpenAI's models. Provide the instruction in the instruction field, and the service processes each source in the batch and returns consolidated results. The OpenAI-hosted equivalent of rag_batch.

Instruction: Enter instruction
Context: Text Source

5.11 anthropic_chat

Sends a prompt to Anthropic's Claude API and returns the response. Provide the instruction in the instruction field, and the service resolves it against the supplied Text Source. Use it when Claude's capabilities are preferred over locally running or other hosted models for a specific node.

Instruction: Enter instruction
Context: Text Source

5.12 gemini_chat

Sends a prompt to Google's Gemini API and returns the response. Provide the instruction in the instruction field, and the service resolves it against the supplied Text Source. Use it when Gemini's multimodal or language capabilities are required for a specific step in the workflow.

Instruction: Enter instruction
Context: Text Source

6. All services

Workspace-specific or user-defined services added for specialized use cases.

Below is the list of supported services, their expected instruction formats, descriptions, and applicable context sources.

Conclusion

With 79 services spanning five categories, Model HQ agents can handle a wide range of tasks — from straightforward document Q&A and text classification, to multi-step data analysis pipelines, multimodal processing, and integrations with external systems and hosted AI providers.

The right approach is to start with the services your workflow actually needs, enable only those, and build up from there. Each service is designed to compose cleanly with the others — the output of one node becomes the input of the next, and the agent state carries results forward through the entire pipeline.