Meet theAI agents

TheAIthatdoestheanalyst’sgroundwork.

Focused agents prepare your data, answer questions and explain results. You see every step.

Data Copilot

Two ways to answer a question.

Pick per conversation. Both use the same scopes and the same guardrails.

Agent mode

The Copilot plans, calls tools and retries until the answer holds up.

  1. list_scopesFinds the Data Scopes that can answer the question.
  2. get_scope_metadataReads tables, columns, descriptions and join rules.
  3. apply_sql_securityValidates scope, adds a row limit, aliases and ordering.
  4. execute_queryRuns the checked SQL at the source.
  5. summarize_conversationKeeps long threads in context.
  6. invoke_mcp_toolCalls a tool on a connected MCP server.

Best for open questions that span several tables or scopes.

Orchestrated mode

A fixed pipeline of specialist agents. Same order, every time.

  1. Query agent
  2. SQL guard
  3. Run at source
  4. Insights agent
  5. Follow-up agent

Best for repeatable questions where speed and consistency matter.

Anatomy of an answer

Seven steps, every one visible.

  1. 01QuestionYou ask in plain English.
  2. 02ScopeThe Copilot picks the Data Scope that can answer it.
  3. 03MetadataIt reads names, types, descriptions and joins. Not your rows.
  4. 04SQLIt writes SQL with real column names.
  5. 05GuardFour checks run before anything touches your data.
  6. 06RunThe query runs in your own database.
  7. 07ExplainAgents explain the result and suggest what to ask next.
Agents

Eleven agents, one job each.

Prepare the data

Agents add the meaning raw schemas lack.

Glossary agent

Writes business definitions for tables and columns.

Reads schema names, types and sample values. Drafts descriptions and glossary terms your team can edit.

Join rules agent

Suggests how tables relate.

Proposes INNER and LEFT joins with the keys to use. You accept or change them in the visual schema builder.

Scope agent

Builds a Data Scope from a goal.

Picks the tables and columns a topic needs, so analysis starts focused.

Scope overview agent

Summarises what a scope holds.

Explains the data, its grain and what it can answer, in plain English.

Answer the question

Agents turn a question into checked SQL and a clear answer.

Data Copilot

Plans the answer and calls tools.

Finds the scope, reads its metadata, writes SQL, runs security checks, executes the query and explains the result. Streams each step.

Query agent

Turns a question into SQL.

Uses only the tables, columns and join rules in your scope. Invented names fail validation.

Insights agent

Explains what the numbers mean.

Calls out trends, gaps and outliers in the result, with the figures behind each point.

Follow-up agent

Suggests the next three questions.

Reads the answer and proposes the analysis most likely to come next.

Starter questions agent

Proposes first questions for a scope.

New users see questions their data can answer, before they type anything.

Work with results

Agents read reports and canvases, not just tables.

Report analysis agent

Answers questions about saved reports.

Chat across one or more reports to compare periods, segments or metrics.

Canvas chat agent

Answers questions from a dashboard.

Reads every card on a canvas, so the team asks without leaving the board.

Models

Your model, your call.

Every agent runs through one AI gateway. Swap models without touching your scopes.

  • Leading families. GPT, Claude, Gemini and Grok.
  • Custom endpoints. Point Anylytics at a model you host.
  • Tuning. Set temperature and token limits per model and per agent.
  • Managed prompts. Each agent's prompt is managed in one place.
  • GPTOpenAI
  • ClaudeAnthropic
  • GeminiGoogle
  • GrokxAI
  • Custom endpointSelf-hosted
  • GPT
  • Claude
  • Gemini
  • Grok
  • Your endpoint
Context

Answers grounded in your own metadata.

Good answers need the right context. Anylytics supplies it on every call.

Semantic search

Catalog assets are embedded in a vector index. Agents find the right table by meaning, not exact name.

Conversation memory

Long threads are summarised, so the tenth question still knows the first.

MCP tools

Connect an MCP server. The Copilot discovers its tools and calls them mid-answer.

What the model sees

  • Your question and the conversation so far
  • Scope metadata: table and column names, types, descriptions, join rules
  • The query result it needs to explain the answer

What it never gets

  • Tables outside the scope you approved
  • Direct access to your database
  • Permission to run SQL that skips the checks

See the agents on your own data.

Free plan, your own data, minutes to the first answer.