Quick answer Website feedback tools like BugHerd, Marker.io, Pastel and Ruttl are still the best way to collect feedback from clients and stakeholders on a live or staging site. As of October 2026, every tool in this guide offers an MCP server, so Claude Code, Codex or Cursor can read their tickets. That leaves a different question: who is pointing, and where? If a client is commenting on a website, keep your feedback tool. If you, the developer, are looking at your own screen and want your agent to fix what you see right now, in any app, a local tool such as Aki (our app, Mac only) is the shorter path.
Disclosure: we make Aki. We've tried to keep this comparison fair, and every claim about another product links to that company's own pages.
What visual feedback tools do
A visual feedback tool (also called a website annotation tool or design feedback tool) lets someone point at a part of a page and leave a comment pinned to it. That beats an email saying "the button on the pricing page looks off."
Most of them work the same way:
- Capture. A reviewer opens the site through a browser extension, an embedded script or widget, or a shared link. Then they click an element or draw on a screenshot and type a comment. BugHerd, for example, offers both a browser extension and a JavaScript snippet, and lets reviewers "drop a pin" on any element (>BugHerd).
- Enrich. The tool attaches metadata on its own. BugHerd records the screenshot, browser, OS, screen resolution, URL and the CSS selector of the element (>BugHerd). Marker.io and Jam also add console logs and network requests (>Marker.io MCP, >Jam docs).
- Route. The feedback becomes a task on a kanban board or in an issue tracker. Marker.io syncs two ways with Jira, GitHub, GitLab, Linear and others (>Marker.io).
These tools were built for a human workflow. A client reports, a project manager triages, and a developer picks up the ticket.
What changes when an AI agent does the fix
When Claude Code or Codex writes the code, the reader of the feedback is a model, not a person. A model works best with text it can act on, not a picture it has to interpret. Here is what a coding agent needs from a piece of visual feedback:
- A precise element reference. A unique CSS selector, or better yet the component source file, so the agent edits the right line instead of guessing from a screenshot.
- The text and HTML of what you pointed at. Then the agent can search the codebase for that string.
- The URL, including the port when it's a local dev server.
- The right repo and session. If you have three worktrees and two agents running, feedback about
localhost:3001should reach the session servinglocalhost:3001. - A small image crop, only when it's needed. A full-page screenshot costs many tokens and is ambiguous. A tight crop helps when the problem is truly visual, like spacing, color or alignment.
- Your note, kept short and next to the element it describes.
Screenshots pasted into a chat cover only items 5 and 6. Classic tickets cover more, and BugHerd's selector capture and Marker.io's console logs are useful context. But the ticket still lives in a cloud workspace, and someone has to hand it to the agent.
The 2026 shift: feedback tools now speak MCP
The good news is that the big tools adapted. All six we checked now offer an MCP (Model Context Protocol) server that lets an agent read their feedback:
- BugHerd has an official MCP server. Agents can read tasks with the screenshot, page URL, CSS selector and browser info, then comment on and update them. BugHerd lists Claude Code, Codex CLI, Cursor, Gemini CLI and others as supported clients (>BugHerd help). It's marked beta and included on all paid plans (>BugHerd pricing).
- Marker.io's MCP server gives agents screenshots, URL, console logs, network requests, browser/OS details and the reporter's description. It's available to all customers at no extra cost (>Marker.io MCP).
- Pastel runs an MCP server at
api.usepastel.com/mcp. For website canvases, the agent is told which page to open, which element the comment is pinned to, and where the screenshot is. It can also reply to comments and update their status (>Pastel help). - Ruttl offers MCP for Claude, ChatGPT, Cursor and Antigravity, so agents can "read feedback, ship fixes, and close threads" (>Ruttl MCP).
- Jam has a hosted MCP endpoint. It loads a recording's video, user events, console logs, errors and network requests into the agent's context, and it's included on the free plan (>Jam docs, >Jam pricing).
- Usersnap's MCP connector is aimed at analyzing customer feedback ("What are customers complaining about in checkout?"), not at handing bugs to a coding agent (>Usersnap help).
So "can my agent read BugHerd?" now has the answer yes. The open question is whether a cloud ticket is the right container for feedback you produce while looking at your own work.
Comparison table
| Tool | Best for | How feedback is captured | Agent access (Oct 2026) | Starting price |
|---|---|---|---|---|
| BugHerd | Client feedback on websites, agencies | Browser extension or JS snippet; pin on element | Official MCP (beta), paid plans | $42/mo annual, $50 monthly; free trial (>pricing) |
| Marker.io | Bug reports from clients and QA into Jira/Linear/GitHub | Widget, browser extension, WordPress plugin; session replay | MCP, all customers | $39/mo annual, $59 monthly; 15-day trial (>pricing) |
| Pastel | Design and website reviews with clients; also images, PDFs | Shareable link, guests comment without login | MCP | Free plan; Pro $35/mo (>plans) |
| Ruttl | Website, mobile app, image and PDF feedback | Comments on live site; NPM package for mobile apps | MCP | Free Basic; Pro $18/user/mo (>pricing) |
| Usersnap | Product feedback at scale (surveys, in-app menu) | In-app widget, annotated screen capture, extension | MCP for feedback analysis | From €49/mo (>pricing) |
| Jam | Bug reports with recordings and dev logs | Chrome extension screen recording / instant replay | MCP, free plan | Free; Team $14/creator/mo annual (>pricing) |
| Aki (ours) | A developer pointing at their own screen for their own agent | Mac app: ⇧⌘A, click an element, drag an area, or hold ⌥ for text, in any app | Local MCP server and aki CLI; routes marks to the right session |
Free to use |
Aki's limits, stated plainly: Mac only (Apple Silicon, macOS 14+). It's local, with no shared workspace, guest links or client portal. Exact element picking needs a Chromium browser with "Allow JavaScript from Apple Events" turned on. Typing the request into the terminal automatically works with the Orca terminal, while other terminals read marks through MCP or the CLI. And it's new (version 0.2.0).
When to keep BugHerd or Marker.io
Aki is not a BugHerd alternative or a Marker.io alternative for the work those tools were built to do. Keep them, or pick one of them, when:
- Clients or stakeholders give the feedback. Non-technical people need a link or widget, a login-free flow and a shared board. Aki runs on one person's Mac.
- Feedback happens on a deployed or staging site across many browsers and devices, and you need the reporter's environment captured.
- You need a record and a workflow. That means assignees, statuses, due dates, and sync with Jira, Linear or ClickUp.
- You want QA context like console logs, network requests and session replays (Marker.io, Jam).
If you're shopping for a BugHerd alternative, the honest shortlist is the other cloud tools. Marker.io leans toward bug reports and issue-tracker sync. Pastel and Ruttl lean toward design review with free entry plans. Jam fits recorded bug reports for developers. All of them can now hand tickets to an agent over MCP. The same goes if you're looking for a Marker.io alternative: compare BugHerd, Jam and Ruttl on price and on which trackers they sync with.
When Aki fits
Aki covers a different moment. You're the developer, an agent is running in your terminal, and you're looking at the result:
- You're reviewing your agent's work on localhost. You click the misaligned card in Chrome. Aki records its unique CSS selector, text, HTML and, in React dev builds, the component's source file. You can use the arrow keys to walk the DOM like DevTools (parent, child, siblings, even
::after). - You run several sessions or worktrees. A page on
localhost:PORTgoes to the session serving that port. Anything else goes to the session you pick in Aki's sidebar. - The thing to fix isn't in a browser. Areas work anywhere, and in other apps (Finder, Xcode, Figma, terminals) Aki uses macOS Accessibility to identify buttons, rows and tabs.
- You care about tokens and privacy. Each mark sends the text it covers first and adds a small crop only when the mark is visual. Marks stay on your Mac, the server listens only on
127.0.0.1, and there's no account or telemetry. The agent you use may still send what it reads to its AI provider.
You queue a few marks, press ⌘⏎, and the agent reads them through Aki's local MCP server (read_annotations, get_annotation_image) and resolves each one.
Many teams will use both: a cloud tool to collect client feedback, and a local tool for the developer's own review loop.
FAQ
What is the best website feedback tool in 2026?
It depends on who gives the feedback. For agencies collecting client comments on websites, BugHerd, Marker.io, Pastel and Ruttl are the mainstream choices. For developer bug reports with logs and recordings, Jam is popular. For product-feedback analysis, Usersnap is built for that.
Can Claude Code or Codex read BugHerd or Marker.io feedback?
Yes. BugHerd's official MCP server lists Claude Code and Codex CLI among its supported clients, and Marker.io's MCP server works with Claude Code, Cursor and other MCP clients (>BugHerd, >Marker.io).
Is Aki a BugHerd alternative?
Only for one use: a developer giving feedback to their own agent. Aki can't collect feedback from clients, has no shared board and runs only on Macs. For client feedback, keep a cloud tool.
Why not just paste a screenshot into the agent?
A screenshot has no selector, text or URL. The agent has to infer which element you mean, and full screenshots use more tokens than a short block of text plus a small crop.
Does Aki work with Safari or Firefox?
Exact element picking (selector, HTML, component file) works in Chromium browsers: Chrome, Brave, Edge, Arc and Vivaldi. In other apps, area marks work everywhere, and Accessibility identifies standard controls.
Is Aki free?
Yes, it's free to use. It's source-available under FSL-1.1-ALv2, and donations are optional. See useaki.vercel.app.
Sources
- BugHerd, homepage: https://bugherd.com/
- BugHerd, pricing: https://bugherd.com/pricing
- BugHerd Help, "Getting started with the BugHerd MCP server": https://support.bugherd.com/en/articles/14835206-getting-started-with-the-bugherd-mcp-server
- Marker.io, homepage: https://marker.io/
- Marker.io, MCP: https://marker.io/marker-mcp
- Marker.io, pricing: https://marker.io/pricing
- Pastel, homepage: https://usepastel.com/
- Pastel, plans: https://usepastel.com/plans
- Pastel Help, "Using the Pastel MCP server": https://help.usepastel.com/en/articles/16399713-using-the-pastel-mcp-server
- Ruttl, homepage: https://www.ruttl.com/
- Ruttl, MCP: https://www.ruttl.com/mcp
- Ruttl, pricing: https://www.ruttl.com/pricing
- Usersnap, homepage: https://usersnap.com/
- Usersnap, pricing: https://usersnap.com/pricing
- Usersnap Help, MCP connector: https://help.usersnap.com/docs/mcp-connector
- Jam, MCP docs: https://jam.dev/docs/debug-a-jam/mcp
- Jam, pricing: https://jam.dev/pricing
- Aki: https://useaki.vercel.app