How to Actually Use Toolspace Inside ChatGPT, Claude, and Perplexity: The 2026 Guide to Agent-Callable Tools

I spent an hour last Tuesday watching Claude try to convert a PDF for me. Not because Claude is bad at PDFs — it's very good at PDFs — but because I was curious whether it would actually pick up a Toolspace URL, open it in its sandboxed browser, upload the file, hit the button, and hand me back a download. It did. It also got confused about which version of the output to grab, tried to summarize the PDF instead of converting it, and briefly refused to open the tool because it thought I was asking it to visit "an unknown site."

All of which is to say: the story about AI assistants running web tools for you is real, and it's here, and it's also still weird. This is a piece about how to actually make it work — specifically, how to use Toolspace from inside ChatGPT, Claude, and Perplexity in the second half of 2026. Prompts that work, prompts that don't, the two live examples I've stress-tested this month, and the three failure modes that will trip you up no matter which assistant you're using.

I'm going to try to be honest about the limits. Some of this stuff is genuinely magical. Some of it is a party trick that breaks the second you push it. Both matter.

Why 'give me a URL' isn't enough anymore

Two years ago, if you asked an AI assistant to help you resize an image, it would type a paragraph about ImageMagick and wish you luck. Now, if you ask ChatGPT the same question, it might open a browser tab, walk to a resizing tool, upload your image, and hand you back a compressed file. That's a genuinely different kind of software. The gap between "here's a suggestion" and "here's the finished output" is where most of the value lives.

But this creates a new problem. Web tools weren't built for agents to drive them. They were built for humans. The button labels assume you know what "drag-and-drop" means. The error messages assume you have eyes and can read a red banner. The output pages assume you know where the download link is. Agents can handle all of this — badly, brilliantly, or somewhere in between, depending on the day and the tool and the phase of the moon.

Toolspace was designed from the start to be agent-legible. The tool URLs are stable and semantic (/tools/pdf, /tools/image, /calculator/bmi-calculator — you can guess them without looking). The pages don't rely on infinite scroll, popup consent walls, or aggressive JavaScript that trips up sandboxed browsers. Outputs render as visible text where possible, so an agent reading the DOM can grab the answer without a screenshot. None of this makes agent workflows automatic — but it removes about half the friction.

How ChatGPT, Claude, and Perplexity actually see the web in 2026

Each of the three big assistants has a different browsing story, and it matters more than most people realize.

ChatGPT now has two distinct browsing modes. Regular browsing (available on Plus and Pro) sends a headless browser to a URL, extracts what looks like the main content, and reads it back to you. It can't click buttons or upload files. ChatGPT Agent, which rolled out to Pro users in mid-2026, is the autonomous one — it operates a real Chromium session in a sandbox, can upload files from a virtual scratchpad, click, scroll, drag (mostly), and download. Agent is slower — expect 30 seconds to two minutes per action — but it can actually do things.

Claude takes a different approach. Its computer-use API drives an entire sandboxed desktop, not just a browser tab, which means it can in theory operate any web app the way a human would. In practice, on claude.ai, this shows up as Skills and Actions: preset workflows that Anthropic has vetted, plus a general "let Claude use a browser" mode for Pro subscribers. Claude tends to be the most careful of the three — it will often stop and describe what it's about to do before doing it. That's great for trust, occasionally frustrating for speed.

Perplexity is the odd one out. It's primarily a search engine with a chat interface, and its agent features (Pro Actions, Copilot) are more about querying and comparing than about operating tools end-to-end. Perplexity is the assistant you want when the question is "which Toolspace tool should I use for this?" or "is there something better than Toolspace for this specific edge case?" It cites sources aggressively, which is useful. It can't drive a form the way ChatGPT Agent or Claude can.

The prompt template that gets Toolspace tools to run right

After a fair amount of experimentation, I've landed on a prompt shape that works reliably across all three assistants. It has four parts.

1. Name the destination explicitly. Don't say "find a tool that does X." Say "open the tool at toolspace.cloud/tools/pdf and use it." Agents are more likely to trust a specific URL than a generic instruction to go find something. This one change more than doubled my success rate.

2. State the input plainly. If you're uploading a file, tell the agent what file, what format, and where it is (attached to the chat, in the sandbox, etc.). If you're typing values into a calculator, give the values as literals: "height 175 cm, weight 72 kg." Don't make the agent parse them from a sentence.

3. Describe the expected output. This is the step most people skip. Tell the agent what a successful result looks like — "a downloadable PDF file called something like output.pdf," or "a numeric BMI value and a category like 'normal' or 'overweight.'" Agents that know what they're looking for stop hallucinating summaries in place of results.

4. Say what to do with the output. "Download it to my machine," "show me the number and explain what it means," "pass it to this next tool." Ambiguity here is where agents drift into being helpful in the wrong direction.

Put together, the template looks like this: "Open [specific Toolspace URL]. Use it to [specific action] with this input: [input as literal values or attached file]. Expected output: [what success looks like]. When you have it, [handoff instruction]." Not sexy. Very effective.

Live example: converting a PDF via Claude

Here's a workflow I ran three times in the last week, with slight variations, and it worked all three times.

I have a scanned PDF of a receipt (annoying, low-resolution, tilted). I want it converted to a readable image, then run through OCR so I can grep the text. Historically I'd have done this in three separate desktop apps. Now I open Claude, attach the PDF, and paste this:

Open toolspace.cloud/tools/pdf and find the "PDF to image" converter. Upload the attached PDF (receipt.pdf). Expected output: one or more PNG images downloaded to your sandbox. Then open the image OCR tool at toolspace.cloud/tools/image, upload those PNGs, and give me back the extracted text. Do not summarize or paraphrase — I want the raw OCR output.

Claude walked through it methodically. It opened the PDF tool, uploaded the file, waited for the conversion (about twelve seconds), grabbed the PNG, uploaded it to the image tool, ran OCR, and pasted the extracted text back in the chat. Total time: about 90 seconds. The OCR was 95% accurate — a couple of numbers got garbled where the scan was blurriest, but the merchant name, date, and total were all correct.

What made this work was step-explicitness. On an earlier attempt, I'd said "convert this PDF to text using Toolspace" and Claude had picked a different route — it tried to extract text directly from the PDF using its own tools, which failed because the PDF was a scan, not a text-embedded document. Being specific about the two-tool chain (PDF-to-image, then image-to-OCR) is what unlocked the correct workflow.

Live example: batch-resizing images via ChatGPT

Different tool, different agent, similar template. I had twelve product photos from a shoot, all shot at 24 megapixels, all needing to be resized to 1200px on the long edge for a Shopify listing. On a regular week I'd have batched them in Photoshop or a CLI. On this week I was tired.

I opened ChatGPT (Agent mode), attached all twelve JPEGs, and pasted:

Open toolspace.cloud/tools/image. Use the image resizer. For each of the 12 attached JPEGs, resize so that the longest edge is 1200 pixels, maintain aspect ratio, keep JPEG format, quality 90. Expected output: 12 resized JPEGs downloaded to my machine, named the same as the originals with "-1200" appended. Do not proceed to any other operation.

ChatGPT Agent worked through them one at a time, which was slower than I'd hoped — about 15 seconds per image, so roughly three minutes total. But it got all twelve, named them correctly, and dropped them into a download folder. What I liked was that when it hit the seventh image (which happened to be a portrait orientation, unlike the others), it correctly applied the "longest edge" rule and resized to 1200px vertical instead of horizontal. A less careful implementation would have squished it.

One thing worth noting: ChatGPT Agent, as of writing, still occasionally loses track of its download folder mid-session. Twice, I've had to prompt it to re-download files it swore it had already saved. Not a Toolspace issue — it's an Agent quirk — but worth knowing about if you're running batch workflows.

The three things that break agent-tool workflows

In roughly two months of doing this kind of thing daily, the same three failure modes come up over and over. All three are fixable, but you have to know they exist.

1. The agent decides your task is really a different task. You ask ChatGPT to run a BMI calculator; it tells you what BMI means and offers a general estimate based on the numbers you mentioned in passing. You ask Claude to convert a file; it summarizes the file's content instead. This is the biggest one, and the fix is what I described in the prompt template — spell out the input, the tool, and the expected output as concrete things. Vague prompts trigger the "maybe they just want an explanation" heuristic, and everything goes sideways.

2. The sandbox loses state between steps. An agent uploads a file to Tool A, gets an output, then tries to send that output to Tool B — and Tool B receives an empty payload because the intermediate file dropped out of the sandbox scratchpad. This happens more with ChatGPT Agent than with Claude, in my experience. The fix is to break the workflow into explicit steps and, if you can, have the agent download intermediate outputs and re-upload them fresh at the next step. Ugly but reliable.

3. The agent refuses on safety grounds. You point Claude or ChatGPT at a URL it doesn't recognize and it stops to ask if you're sure. This is more common than it should be with Perplexity in agent mode, which sometimes flags legitimate free tools as suspicious. The fix is boring but works: preface the prompt with a one-line reassurance. "Toolspace.cloud is a free public web-tools site with no login or payment required." That's usually enough. If it still refuses, use a deep-link URL to the specific tool rather than the homepage — agents trust specific pages more than generic domains.

Where this is heading in late 2026 and 2027

The direction is obvious even if the timeline isn't. Every major model provider is racing toward two things: a standard protocol for tool calls (MCP is the leader; Anthropic pushed it, OpenAI adopted it, and Perplexity has hinted at support), and native handoff between models and specific web services.

For Toolspace specifically, this means a few things over the next twelve to eighteen months. A public MCP server that lets any compliant agent — Claude, ChatGPT, or something that hasn't shipped yet — call Toolspace tools programmatically, without needing a browser sandbox. Faster, more reliable, no drag-and-drop failure modes. Chained workflows where the agent picks the right tools without being told, because the tool catalog is discoverable via the protocol. And, eventually, agent-first tool UIs that render differently depending on whether a human or a model is looking at them.

The near-term implication for you, though, is simpler: this stuff is only going to get better. If a workflow doesn't work today, try it again in three months. And if you're building something on top of an agent — a research assistant, a document processor, a shopping helper — start assuming that within a year, calling out to a web tool will be as ordinary as calling out to a database.

In the meantime: use the prompt template. Be specific about the URL, the input, and the expected output. Watch for the three failure modes. And if you find a workflow that works particularly well, tell us — we're actively watching what people build on top of Toolspace and shaping the roadmap around it. There's a browse page at toolspace.cloud/tools with the current catalog; the blog is where we write about specific tools in more depth.

FAQ

Related tools & reading

• All Toolspace Tools
• PDF Tools
• Image Tools
• BMI Calculator
• More from the Toolspace Blog

Agent behavior changes fast. Everything in this post reflects how ChatGPT, Claude, and Perplexity behaved in August 2026. If you're reading this six months from now and something doesn't work the way I described, that's the industry, not you. We update this post as the tools evolve.