Evidence-first authority page
AI agent glossary: agents, frameworks, models, MCP and evidence
Plain-language definitions for the entity and evidence terms used across BestAIAgent.in.
Direct answer
BestAIAgent.in uses distinct terms for agents, frameworks, models, providers and evidence so that claims do not leak from one entity type to another. The glossary explains those boundaries and links each concept to its canonical directory.
Last reviewed 2026-08-23
This page remains noindex while its page-level source receipts are completed and verified. It is available for review, not presented as publication-ready authority.
Agent
A software product or project that uses one or more models to pursue a task through a workflow that can include reasoning, tools, memory, retrieval or execution.
Agent framework
A developer-oriented library or orchestration system used to build, coordinate or control agent workflows. A framework is not automatically an end-user agent product.
Model and provider
A model is the inference artifact or service performing model computation. A provider is the organization or service publishing models or related infrastructure. One provider can publish multiple models, and one agent can use multiple providers.
MCP
Model Context Protocol is an open protocol used to connect AI applications with external context and tools. BestAIAgent.in treats MCP server identity and MCP server capability claims as separately verifiable fields.
Evidence receipt
A stored record containing the source URL, publisher, retrieval date, content hash and a normalized snapshot of the fact being relied on. Evidence receipts support auditability but do not make unrelated fields verified.
LLM (large language model)
A neural network trained on large text corpora to predict and generate language. Most AI agents on this site are built on top of one or more LLMs accessed via an API or run locally.
Context window
The maximum amount of text (measured in tokens) a model can process in a single request, including the prompt, conversation history and any retrieved documents.
Token
A unit of text — roughly a word fragment — that a model reads and generates. Pricing and context-window limits are typically expressed in tokens rather than words or characters.
RAG (retrieval-augmented generation)
A pattern where a model's response is grounded by retrieving relevant documents from an external index (often a vector database) before generating an answer, rather than relying solely on parameters learned during training.
Fine-tuning
Further training a pretrained model on a narrower, task-specific dataset to adapt its behavior, distinct from prompting, which shapes behavior only at inference time.
Tool use / function calling
A model capability where the model can request execution of an external function or API — for example a search, code-execution or database-query tool — and incorporate the result into its response.
Hallucination
A model output that is fluent and confident but factually incorrect or unsupported by any source. It is the primary reason this site labels fields unknown rather than inferring them from a model's own output.
Inference
The process of running a trained model on new input to produce an output, as opposed to training, which is the process of learning the model's parameters.
Open-weight model
A model whose trained parameters are published for download and local use, as distinct from a closed model that is only accessible through a hosted API.
Sandboxed execution
Running an agent's code, shell commands or tool calls inside an isolated environment (such as a container) so that actions cannot affect the host system without explicit permission.