Wield vs Integuru: what each does and when to pick it
Integuru and Wield both treat a website's network requests as the interface, but they differ in what you get. Integuru v0 records a HAR file and generates integration code that you run yourself. Wield ships signed JSON recipes that run in your own browser, with no model in the loop. Pick Integuru when you want code you own. Pick Wield when credentials must stay on your device.
What Integuru is
Integuru is an open-source project licensed under AGPL-3.0. Its GitHub README describes the v0 method: capture the browser's network requests, then generate runnable code for a platform that has no official API.
The method has four steps. You capture a HAR file with create_har.py, save your cookies to cookies.json, and describe the action in a prompt. The agent then finds the request that performs the action and its dynamic parameters. It builds a dependency graph until only the auth cookies are needed, then generates code from the leaf requests up to the main request.
The README says the project has moved on to integuru.com. This page covers only the v0 behavior that the README documents. For the current product, check the project's own site.
Side-by-side
| Wield | Integuru (v0) | |
|---|---|---|
| What you get | Recipes: public, signed JSON files that describe the requests a site's pages make | Runnable integration code, generated for each task |
| Where it runs | Your own browser, through the Wield extension | Code you run on your own servers |
| How a site is captured | The Wield recorder captures one task's network calls and strips cookies, auth headers, tokens and body values on your device | create_har.py captures a HAR file; cookies go to cookies.json |
| Signing | Every recipe is signed with Ed25519; unsigned or changed recipes are refused | Not covered in the README |
| Model in the loop | None when a recipe runs, so a run uses no LLM tokens | Cloud LLMs (GPT-4o and o1-preview, per the README) build the code |
| License | MIT planned for the executor, bridge, MCP server and recipe format; source not public yet | AGPL-3.0 |
| Cookies | Wield never reads, copies or sends them | Saved to cookies.json for the generated code |
Logins and your data
With Integuru, you log in yourself during capture and complete any 2FA by hand. The README says the HAR file and cookies are stored locally, but the data is sent to cloud LLMs. Read the README before you capture a logged-in session.
With Wield, recipes run in your browser through the extension, using the session you already have. Passwords and cookies never leave your computer. Results go from the extension to wield-mcp and your AI app, and never pass through wield.run. The recorder strips secrets on the device before the hosted recipe builder sees the trace. Recording and the builder are paid-plan features.
When Integuru is the better choice
- You want code you own, which you can read, change and deploy on your own servers.
- You are comfortable exporting cookies to a file and running code that uses them.
- You want an AGPL-3.0 project that you can inspect and modify.
- You need one integration that you maintain yourself, rather than a lookup that an AI app runs for you.
When Wield is the better choice
- You want an AI app such as Claude, Cursor or Codex to call a site as a tool, from your own browser.
- Credentials must not leave the device, and cookies must never be read or copied.
- You want runs that use no model and no integration code that you maintain.
- You want writes to need your approval. Wield shows every write request before it leaves your browser, and Deny is the default.
Install Wield and read the quick start to try it.
FAQ
Is Wield a replacement for Integuru?
Not exactly. Both use browser network requests as the interface. Integuru produces code that you deploy. Wield runs recipes that an AI app calls. Choose by where the credentials must live and who runs the integration.
Does Wield generate code?
No. A Wield recipe is a small JSON file that describes requests. Your agent calls it through wield-mcp, and nothing is generated for each task.
Which one costs less to run?
A Wield recipe run uses no model, so it uses no LLM tokens. Integuru's README names cloud LLMs in its workflow, so model use is part of its design. Check the Integuru project's pricing page for current costs.
What data does Integuru send to cloud LLMs?
The README says the HAR file and cookies are stored locally, while the data is sent to cloud LLMs, named as GPT-4o and o1-preview. Read the README for the current details before you capture a logged-in session.
Sources
- Integuru on GitHubgithub.com