WorkGPT is a GPT agent framework for calling APIs, with a design approach similar to AutoGPT or LangChain. The user provides the framework with a directive and a set of APIs, and the framework engages in multiple rounds of conversation with the AI until the directive is completed. For example, the directive can be to research something on the web, scrape a website, or order an Uber for the user.
In terms of main capabilities, WorkGPT supports any API that can be represented with an OpenAPI file. It does not force the use of OpenPM (a package manager for OpenAPI files), and also supports importing any OpenAPI file. For authentication handling, the user only needs to pass in authKey, and the library automatically determines how to authorize. All endpoints in the API are exposed to the LLM as local functions for it to call.
In terms of usage, the user needs to provide a directive and an array of APIs, and the framework drives the AI to interact with the APIs until the directive is completed. The example shows pulling a package named ipinfo from OpenPM to query IP addresses. In addition, the project includes an example of using Puppeteer as a text browser to enable the LLM to access the web. This text browser outputs all HTML and returns only text, which is already sufficient for GPT-4 because it can extract data from plain text. Users can also pass a custom API to the LLM as a "completion program API", to be called when the LLM completes, so that a schema to follow can be provided, making it easier to extract data from web pages.
Applicable scenarios include: needing AI to automatically call external APIs to complete tasks such as research, scraping, and ordering; needing to quickly connect any OpenAPI service to an LLM agent; needing to let the LLM access web pages through a text browser and extract information.
In terms of limitations, the material does not provide detailed information about charging, pricing, platform support, performance results, or specific usage steps; this content is not yet provided.
