All tool executions now return error objects instead of throwing exceptions. This allows the AI model to see errors and respond appropriately instead of causing the entire stream to fail silently. |
||
|---|---|---|
| .. | ||
| src | ||
| CHANGELOG.md | ||
| package.json | ||
| README.md | ||
| tsconfig.json | ||
@tpmjs/registry-execute
Execute tools from the TPMJS registry in any AI SDK agent. Tools run in a secure sandbox - no local installation required.
Installation
npm install @tpmjs/registry-execute
# or
pnpm add @tpmjs/registry-execute
Usage
import { streamText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { registrySearchTool } from '@tpmjs/registry-search';
import { registryExecuteTool } from '@tpmjs/registry-execute';
const result = streamText({
model: anthropic('claude-sonnet-4-20250514'),
tools: {
registrySearch: registrySearchTool,
registryExecute: registryExecuteTool,
},
system: `You have access to the TPMJS tool registry.
Use registrySearch to find tools, then registryExecute to run them.`,
prompt: 'Search for web scraping tools and scrape https://example.com',
});
// The agent can now:
// 1. Search for tools: registrySearch({ query: "web scraping" })
// 2. Execute found tools: registryExecute({ toolId: "@firecrawl/ai-sdk::scrapeTool", params: { url: "..." } })
Tool: registryExecuteTool
Execute a tool from the TPMJS registry by its toolId.
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
toolId |
string | Yes | Tool identifier (format: package::name) |
params |
object | Yes | Parameters to pass to the tool |
env |
object | No | Environment variables (API keys) if required |
Example
// Execute a web search tool
const result = await registryExecuteTool.execute({
toolId: '@exalabs/ai-sdk::webSearch',
params: { query: 'latest AI news' },
env: { EXA_API_KEY: 'your-api-key' },
});
// Result:
// {
// toolId: '@exalabs/ai-sdk::webSearch',
// executionTimeMs: 1234,
// output: { results: [...] }
// }
Returns
{
"toolId": "@exalabs/ai-sdk::webSearch",
"executionTimeMs": 1234,
"output": { ... }
}
Environment Variables
| Variable | Default | Description |
|---|---|---|
TPMJS_API_URL |
https://tpmjs.com |
Base URL for the registry API |
TPMJS_EXECUTOR_URL |
https://executor.tpmjs.com |
URL for the sandbox executor |
Self-Hosted Registry
To use your own TPMJS registry and executor:
export TPMJS_API_URL=https://registry.mycompany.com
export TPMJS_EXECUTOR_URL=https://executor.mycompany.com
Passing API Keys to Tools
Many tools require API keys to function (e.g., web scraping tools need a Firecrawl key, search tools need an Exa key). The recommended approach is to wrap registryExecuteTool with your pre-configured keys.
Recommended: Create a Wrapper
import { tool } from 'ai';
import { registryExecuteTool } from '@tpmjs/registry-execute';
// Pre-configure your API keys
const API_KEYS: Record<string, string> = {
FIRECRAWL_API_KEY: process.env.FIRECRAWL_API_KEY!,
EXA_API_KEY: process.env.EXA_API_KEY!,
};
// Create a wrapped version that auto-injects keys
export const registryExecute = tool({
description: registryExecuteTool.description,
parameters: registryExecuteTool.parameters,
execute: async ({ toolId, params }) => {
return registryExecuteTool.execute({ toolId, params, env: API_KEYS });
},
});
Now use the wrapped tool in your agent:
import { streamText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { registrySearchTool } from '@tpmjs/registry-search';
import { registryExecute } from './tools'; // Your wrapped version
const result = streamText({
model: anthropic('claude-sonnet-4-20250514'),
tools: {
registrySearch: registrySearchTool,
registryExecute, // Keys are auto-injected
},
system: `You have access to the TPMJS tool registry.
Use registrySearch to find tools, then registryExecute to run them.`,
prompt: 'Scrape https://example.com and summarize the content',
});
How It Works
-
Search returns required keys: When you search for a tool, the response includes
requiredEnvVars:{ "toolId": "@firecrawl/ai-sdk::scrapeTool", "requiredEnvVars": ["FIRECRAWL_API_KEY"] } -
Your wrapper injects keys: The wrapped tool automatically passes your configured keys to the executor.
-
Keys are injected into sandbox: The executor injects these as environment variables in the isolated Deno runtime where the tool runs.
Tools Without Required Keys
Some tools don't require any API keys (like @tpmjs/createblogpost). They work with or without the wrapper.
Security
- All tools run in a sandboxed Deno environment on Railway
- API keys are passed per-request, never stored
- Keys are isolated to the specific tool execution
- Only registered tools can be executed (no arbitrary code)
Related
- @tpmjs/registry-search - Find tools to execute
- TPMJS Registry - Browse all available tools
License
MIT