Tools Overview

DB-GPT ships with a small set of built-in tools that power the Agentic Data API. These tools are the default execution surface for data analysis, skill-driven workflows, SQL exploration, shell access, and HTML report delivery.

The current source of truth lives in:

packages/dbgpt-app/src/dbgpt_app/openapi/api_v1/agentic_data_api.py

Built-in tools

The core built-in tools are:

load_skill
code_interpreter
shell_interpreter
sql_query
html_interpreter

They are exposed as agent tools and used by the Agentic Data workflow to move from reasoning → execution → presentation.

Tool selection guide

Tool	Use it for	Do not use it for
`load_skill`	Load a skill's instructions and workflow	Running code or shell commands
`code_interpreter`	Python analysis, charts, dataframe logic, calculations	Shell commands or final HTML rendering
`shell_interpreter`	Bash / CLI commands such as `ls`, `grep`, `curl`, `git`, `pip`	Python analysis or skill-specific script execution when a skill says otherwise
`sql_query`	Read-only SQL against the selected datasource	Any write/update/delete schema changes
`html_interpreter`	Final HTML page / report rendering	General Python computation or shell execution

Typical execution flow

For most agentic data tasks, the pattern is:

Use load_skill when the task matches a reusable skill.
Use sql_query to inspect structured data.
Use code_interpreter for Python analysis, chart generation, and data shaping.
Use shell_interpreter only for true shell / CLI work.
Use html_interpreter as the final presentation step for reports or web pages.

Important rules

1. `html_interpreter` is the final presentation tool

When a user asks for:

an HTML report
a web page
an interactive report
a rendered analysis deliverable

the final rendering step should go through html_interpreter.

2. `sql_query` is read-only

sql_query only supports safe query access. It is designed for SELECT-style exploration, not data mutation.

3. `code_interpreter` calls are independent

Each code_interpreter call runs independently. Variables do not persist between calls, so every snippet must include its own imports, loading logic, and output statements.

4. `shell_interpreter` is for shell tasks only

Use shell_interpreter for CLI workflows. If a skill specifically requires another execution path, follow the skill instructions.

Next step

See Tool Resource for the detailed meaning, parameters, examples, and usage patterns of each built-in tool.

Source: docs/i18n/zh-CN/docusaurus-plugin-content-docs/current/agents/introduction/tools.md