System Tool Development Guide
Learn how to build FastGPT system tool plugins from scratch, define input/output schemas and secrets, implement handlers, and package releases.
Introduction
This document targets system tool development after FastGPT v4.15.0. System tools use SDK definitions and are delivered to the FastGPT Plugin service as .pkg files for installation, updates, and runtime management. Model presets use a static configuration contribution workflow.
The currently stable system tool plugin types are:
- Single tool: one plugin exposes one tool and is declared with
defineTool(). - Toolkit: one plugin exposes multiple related child tools and is declared with
defineToolSet().
System tool plugins run in the runtime provided by the FastGPT Plugin service. The FastGPT main service invokes tools through the plugin service, and plugin code uses @fastgpt-plugin/sdk-factory to describe input, output, secret configuration, and execution logic.
Differences From The Legacy Mechanism
- FastGPT and FastGPT Plugin remain separate services within a microservice architecture.
- The plugin package protocol upgrades from the old built-in system tool directory to a unified
.pkgformat, making installation, version management, hot updates, and future plugin type expansion easier. - The plugin runtime is managed by the server. The current default runtime is
local-pool, where each plugin version has its own process pool, queue, and runtime configuration. - Plugin metadata, input/output schemas, secret schemas, and icon assets are included in build artifacts for use by FastGPT pages, workflows, and Agents.
- Tool development uses
@fastgpt-plugin/cliand@fastgpt-plugin/sdk-factoryto define schemas and build packages.
Information To Collect Before Development
Clarify these items before coding:
| Information | Description |
|---|---|
| Plugin type | tool or tool-suite. |
| Plugin ID | pluginId, globally stable and unique. Keep it unchanged after publishing. |
| Child tool ID | Required for Toolkits. children[].id stays unchanged after publishing. |
| Chinese and English names | name.en and name.zh-CN. |
| Chinese and English descriptions | description.en and description.zh-CN. |
| Inputs | Type, constraints, default value, UI title, and description for each field. |
| Outputs | Type, meaning, and downstream usage for each field. |
| Secrets | API Key, Base URL, username/password, and similar values, described through secretSchema. |
| External API | Request method, auth method, timeout, rate limit, error response, and test account. |
| File capability | Use ctx.invoke.uploadFile() when file upload is needed. |
| Streaming output | Use ctx.streamResponse() when streaming text answers to the user. |
| Test cases | Include at least success, invalid parameters, auth failure, and upstream failure. |
Missing information that affects plugin ID, auth method, billing, or listing security should be confirmed first. Other missing information can use reasonable defaults, with assumptions recorded in the submission notes.
Developing With An Agent
When using Claude Code, Codex, or another agent tool, copy this prompt:
Develop a plugin using the following official FastGPT plugin development Skill:
https://raw.githubusercontent.com/labring/fastgpt-official-plugins/refs/heads/main/.agents/skills/develop-fastgpt-plugin/SKILL.md
Requirements:
1. Read and understand the complete Skill first, then follow its development workflow.
2. Before coding, collect the plugin name and type, Chinese and English names and descriptions, inputs and outputs, secrets, external APIs, expected behavior, error handling, and test cases.
3. If key requirements are missing, ask no more than three essential questions. When reasonable defaults are available, state the assumptions and continue.
4. Use `@fastgpt-plugin/cli` to create the plugin skeleton. Follow the repository's existing structure, naming, testing, and build conventions.
5. After implementation, run the required validation, including tests, build, plugin checks, and packaging. Explain anything that could not be validated.
6. In the final response, list changed files, validation results, remaining assumptions, and external API behavior that requires human confirmation.When developing or maintaining SDK/CLI in the fastgpt-plugin repository, also refer to local skills:
sdk/factory/skills/fastgpt-plugin-development/SKILL.mdsdk/factory/skills/fastgpt-system-tool-development/SKILL.mdsdk/factory/skills/fastgpt-sdk-factory/SKILL.md
1. Prepare Environment
Recommended environment:
- Node.js 22 (
>= 22.23.2). pnpm10; thefastgpt-pluginrepository uses pnpm workspace.- Git.
- GitHub CLI
gh, used for forking, creating repositories, and submitting PRs.
When developing community plugins, first fork and clone the community repository:
gh repo fork labring/fastgpt-community-plugins --clone
cd fastgpt-community-plugins
pnpm installWhen debugging the CLI or SDK in the fastgpt-plugin repository, install dependencies and build the CLI/SDK first:
pnpm install
pnpm build:sdk-factory
pnpm build:cli2. Create Plugin Skeleton
Single-tool plugin:
pnpx @fastgpt-plugin/cli create my-tool --type tool --cwd packages/toolsToolkit plugin:
pnpx @fastgpt-plugin/cli create my-tool-suite --type tool-suite --cwd packages/toolsYou can also enter the target directory and create interactively:
pnpx @fastgpt-plugin/cli createThe CLI creates the plugin directory and common files:
| File | Purpose |
|---|---|
index.ts | Plugin entry, default-exporting defineTool() or defineToolSet(). |
index.spec.ts | Unit test file. |
package.json | Plugin dependencies and build, build:dev, check, dev, debug, debug:run, pack, and test scripts. |
tsconfig.json | TypeScript config. |
vitest.config.ts | Test config. |
README.md | Default plugin documentation. |
README.zh-CN.md | Simplified Chinese plugin documentation. |
README.en.md | English plugin documentation. |
logo.svg | Main plugin icon. |
3. Implement Single Tool
The system tool entry must default-export an SDK factory instance:
import {
createToolHandler,
defineTool,
type InputSchemaMetaType,
type OutputSchemaMetaType,
type SecretSchemaMetaType
} from '@fastgpt-plugin/sdk-factory';
import z from 'zod';
const secretSchema = z.object({
apiKey: z
.string()
.min(1)
.meta({
title: 'API Key',
isSecret: true
} satisfies SecretSchemaMetaType)
});
const handler = createToolHandler({
inputSchema: z.object({
query: z
.string()
.min(1)
.meta({
title: 'Query',
description: 'Search keyword'
} satisfies InputSchemaMetaType)
}),
outputSchema: z.object({
result: z.string().meta({
title: 'Result'
} satisfies OutputSchemaMetaType)
}),
secretSchema,
handler: async (input, ctx) => {
return {
result: input.query
};
}
});
export default defineTool({
manifest: {
pluginId: 'example-search',
version: '1.0.0',
name: {
en: 'Example Search',
'zh-CN': '示例搜索'
},
description: {
en: 'Search example data',
'zh-CN': '搜索示例数据'
},
versionDescription: {
en: 'Initial version',
'zh-CN': '初始版本'
},
tags: ['tools']
},
handler
});Core rules:
- Keep
pluginId, child toolid, input field names, and output field names stable after publishing. - Use
{ en, 'zh-CN' }formanifest.name,manifest.description, andversionDescription. - Describe inputs, outputs, and secrets with Zod schemas.
- Add
InputSchemaMetaTypeto input fields andOutputSchemaMetaTypeto output fields. - Add
SecretSchemaMetaTypeto secret fields and setisSecret: truefor sensitive fields. - Handler return values must match
outputSchema. - Convert external API errors into actionable messages and avoid exposing secrets, tokens, or complete sensitive responses.
- Use
ctx.invoke.uploadFile()when host file upload is needed, and prefer preserving the returnederr. - Use
ctx.streamResponse({ type: 'answer' | 'fastAnswer', content })when streaming text answers to users.
4. Implement a Toolkit
Use defineToolSet() for Toolkits. Put shared information in the top-level manifest and secretSchema, and declare each child tool's independent id, name, description, and handler in children.
import {
createToolHandler,
defineToolSet,
type InputSchemaMetaType,
type OutputSchemaMetaType,
type SecretSchemaMetaType
} from '@fastgpt-plugin/sdk-factory';
import z from 'zod';
const secretSchema = z.object({
apiKey: z.string().meta({
title: 'API Key',
isSecret: true
} satisfies SecretSchemaMetaType)
});
const searchHandler = createToolHandler({
inputSchema: z.object({
query: z.string().meta({
title: 'Query'
} satisfies InputSchemaMetaType)
}),
outputSchema: z.object({
items: z.array(z.string()).meta({
title: 'Items'
} satisfies OutputSchemaMetaType)
}),
secretSchema,
handler: async (input) => ({ items: [input.query] })
});
const summaryHandler = createToolHandler({
inputSchema: z.object({
content: z.string().meta({
title: 'Content'
} satisfies InputSchemaMetaType)
}),
outputSchema: z.object({
summary: z.string().meta({
title: 'Summary'
} satisfies OutputSchemaMetaType)
}),
secretSchema,
handler: async (input) => ({ summary: input.content.slice(0, 100) })
});
export default defineToolSet({
manifest: {
pluginId: 'text-tools',
version: '1.0.0',
name: {
en: 'Text Tools',
'zh-CN': '文本工具集'
},
description: {
en: 'Search and summarize text',
'zh-CN': '搜索和总结文本'
}
},
children: [
{
id: 'search',
name: { en: 'Search', 'zh-CN': '搜索' },
description: { en: 'Search text', 'zh-CN': '搜索文本' },
handler: searchHandler
},
{
id: 'summary',
name: { en: 'Summary', 'zh-CN': '总结' },
description: { en: 'Summarize text', 'zh-CN': '总结文本' },
handler: summaryHandler
}
],
secretSchema
});5. Icon Conventions
During build, the CLI scans icons in the plugin root and writes them into the built manifest.json.
| Scenario | File name |
|---|---|
| Main plugin icon | logo.svg, logo.png, logo.jpg, logo.jpeg, logo.webp, or logo.gif |
| Toolkit child icon | <childId>.logo.svg, <childId>.logo.png, and similar names |
Notes:
- Put icon files in the plugin root.
- The
<childId>of a child icon must exactly matchchildren[].id. - Keep only one extension for the same icon to avoid ambiguous scan results.
- Child tools without their own icons reuse the main plugin icon by default.
- After build, check the
iconfield indist/manifest.json.
6. Key Development Recipes
6.1 Authenticated External API Requests
When calling external third-party services requiring an API key, declare the secret field in secretSchema and extract it via ctx.secrets in the handler. Handle network and HTTP errors with actionable messages while keeping credentials omitted from logs and error text:
import {
createToolHandler,
defineTool,
type InputSchemaMetaType,
type OutputSchemaMetaType,
type SecretSchemaMetaType
} from '@fastgpt-plugin/sdk-factory';
import z from 'zod';
const secretSchema = z.object({
apiKey: z
.string()
.min(1)
.meta({
title: 'API Key',
isSecret: true
} satisfies SecretSchemaMetaType)
});
const UPSTREAM_SEARCH_URL = 'https://api.example.com/v1/search';
const handler = createToolHandler({
inputSchema: z.object({
query: z
.string()
.min(1)
.meta({
title: 'Query',
description: 'Search keyword sent to the trusted upstream service'
} satisfies InputSchemaMetaType)
}),
outputSchema: z.object({
status: z.string().meta({
title: 'Status'
} satisfies OutputSchemaMetaType),
data: z.any().meta({
title: 'Response Data'
} satisfies OutputSchemaMetaType)
}),
secretSchema,
handler: async (input, ctx) => {
try {
const requestUrl = new URL(UPSTREAM_SEARCH_URL);
requestUrl.searchParams.set('q', input.query);
const res = await fetch(requestUrl.toString(), {
headers: {
Authorization: `Bearer ${ctx.secrets.apiKey}`,
'Content-Type': 'application/json'
}
});
if (!res.ok) {
throw new Error(`Upstream service returned status ${res.status}`);
}
const data = await res.json();
return { status: 'success', data };
} catch (err: any) {
throw new Error(`External API request failed: ${err.message || 'unknown error'}`);
}
}
});Keep the authenticated upstream endpoint fixed in a trusted server constant or controlled configuration so ctx.secrets credentials are sent only to verified hosts.
6.2 Host File Uploads And Artifact Return
When a tool produces generated files (such as reports, exported images, or media) and returns them to FastGPT callers, invoke the host upload capability ctx.invoke.uploadFile(). The host stores the artifact and returns an accessible URL:
handler: async (input, ctx) => {
const content = Buffer.from(
`Report for ${input.query}\nGenerated at ${new Date().toISOString()}`
);
const [result, err] = await ctx.invoke.uploadFile({
file: content,
fileName: 'report.txt',
contentType: 'text/plain'
});
if (err) {
throw new Error(`Failed to upload generated file: ${err}`);
}
return {
fileUrl: result.accessURL,
filename: 'report.txt'
};
};6.3 Streaming Text Answer Output
To stream text answer chunks directly to the conversation client during execution, call ctx.streamResponse() with { type: 'answer' | 'fastAnswer', content: string }:
handler: async (input, ctx) => {
await ctx.streamResponse({
type: 'answer',
content: 'Fetching remote resources...\n'
});
// Execute processing logic
await ctx.streamResponse({
type: 'answer',
content: 'Processing complete.\n'
});
return { result: 'Operation completed successfully' };
};6.4 Side Effects And Repeated Invocation Handling
FastGPT Workflows and Agents may call a tool multiple times during loop executions or retries. For operations with side effects (such as database writes, external charges, or messaging):
- Business idempotency: Include a unique identifier (such as
requestIdororderId) in the input schema, and enforce idempotency on the external service. - State awareness: Return existing outcomes for already completed external actions.
- Explicit timeouts: Set bounded timeouts on requests, detect cancellation signals, and stop pending downstream work promptly.
7. Debugging, Packaging, And Publishing Workflow
After implementing your tool, follow the unified delivery path in debugging, packing, and publishing:
- Local debugging: Use
pnpx @fastgpt-plugin/cli debug . --runto verify schemas, defaults, and boundary inputs. See local debugging. - Remote debugging: Connect to your FastGPT test environment using
pnpx @fastgpt-plugin/cli dev. See remote debugging. - Build and pack: Run
pnpm run build,cli check, andpnpm run packto generate.pkgartifacts. See build, check, and pack. - Pre-submission verification: Check items against the verification checklist.
- Publishing channels: Choose community distribution, official distribution, or private commercial delivery. See publishing workflow.
Design and Versioning Guidelines
How should I choose between tool and tool-suite?
Use tool for a single capability. Use tool-suite for multiple capabilities that share authentication, the same upstream API, and strong business relevance, such as search, detail, and task creation in one plugin.
How should plugin versions be managed?
Use semantic versioning for manifest.version. Upgrade patch for compatible fixes, minor for compatible new features, and major when changing input/output fields, child tool IDs, or user configuration. Evaluate existing workflow compatibility before major changes.
Can I put API keys in code or environment variables?
Plugins should declare secrets through secretSchema and read them through ctx.secrets. Real secrets should not appear in code repositories, test snapshots, error logs, or README files.
Is a test environment still needed after local debug passes?
Yes. Local debug quickly validates plugin logic and schemas. Test environment validation confirms real installation, runtime, host reverse invocation, network, and permission behavior.
References
Last updated