magnacoda
Welcome
magnacoda™ is an accessibility scanner for WCAG, the Web Content Accessibility Guidelines, that you run from your terminal. It's built for developers and content managers who want to understand and fix accessibility problems, not just list them.
These are screens from magnacoda, from a scan of a sample page.
Swipe sideways to see the whole screen.
Scan
magnacoda » scan https://unrealbakeshop.example/
Ctrl+C or q cancels at any time.
The defaults: WCAG 2.2 AA, a desktop screen, light and dark mode, no AI.
You can type: options (such as --option), defaults, help. Enter cancels.
Options » defaults
Scanning https://unrealbakeshop.example/ - 5 s so far
A scan of the sample page, from magnacoda in a terminal.
Results
magnacoda » scan https://unrealbakeshop.example/
Ctrl+C or q cancels at any time.
The defaults: WCAG 2.2 AA, a desktop screen, light and dark mode, no AI.
You can type: options (such as --option), defaults, help. Enter cancels.
Options » defaults
Tested on desktop, in light mode (the page looks the same in dark mode).
Found 12 problems on 12 elements (2 critical, 9 serious, 1 moderate), and
49 items to check.
The scan took 19 s.
Saved in ~/Documents/magnacoda-scans/unrealbakeshop.example_20261119-105414.
Type review to see them in your browser, and recheck once
you've fixed some.
magnacoda »
The scan's results, from magnacoda in a terminal.
Review
9. input ("Your email")
- Fix 1.3.5 Identify Input Purpose (opens in a new tab) (AA) - It looks like it asks for the visitor's own email address, but has no autocomplete value, so browsers can't fill it in and assistive tools can't show what it's for (moderate; rule magnacoda/missing-autocomplete)
- now:
<input type="email" placeholder="Your email"> - fix: Add
autocomplete="email", if the field asks for the visitor's own details (for someone else's, such as a gift's recipient, leave it out). - e.g.
<input type="email" placeholder="Your email" autocomplete="email">
- now:
- Checked: tick when you've looked at everything below that needs a person
- Check 3.3.1 Error Identification (opens in a new tab) (A) - Submit invalid or empty input: errors must be shown in text and identify the field in error.
- Check 3.3.3 Error Suggestion (opens in a new tab) (AA) - When an input error is detected and a fix is known, the message should suggest it.
- Check 4.1.3 Status Messages (opens in a new tab) (AA) - Status messages (success, errors, result counts) must be announced via role=status/alert or live regions without moving focus.
One finding from the scan of the sample page, on the review page in your browser.
Help
magnacoda - Help
magnacoda is a WCAG 2 accessibility scanner for developers and content
managers comfortable in a terminal. It explains what it finds and how to fix
it, so you can act on the results right away and learn as you go.
- Open a terminal and run magnacoda with no arguments to start an
interactive shell. Let the wizard ask a few plain-language questions and
build the scan for you. To learn how to open a terminal on your computer,
type docs starting.
- Or scan a page directly from the command line with magnacoda scan URL
[OPTIONS]. The URL can be short, such as example.com or localhost:3000:
magnacoda fills in the rest.
- To use the AI tiers you can provide your own Anthropic API key (setup
can save one for you), or use Claude Code if it is installed and logged
into your Claude subscription.
- AI works with magnacoda both ways: your AI coding agent can run it, and
it can use AI itself. To see both in one round of work on a page, type
The help, from magnacoda in a terminal. You can read all of it in Read the docs.
Setup
magnacoda » setup
Ctrl+C or q cancels at any time.
Chromium is already installed.
A shortcut opens magnacoda in a terminal window, from your app menu or
desktop.
Add a magnacoda shortcut to (menu, desktop, both, skip) [skip] »
AI tiers (optional; magnacoda works fine without them):
Claude Code is installed, so the AI tiers can use your Claude plan (such as
Pro or Max) instead of an Anthropic API key, which the API bills per use.
Use your Claude plan for the AI tiers? [Y/n] »
Setup, from magnacoda in a terminal.
magnacoda docs
Manual: AI tiers
AI and automation
AI is optional and off until you turn it on; AI tiers explains what each level adds. The rest is for running magnacoda with no one at the keyboard: in a CI pipeline, a script or an AI coding agent.
--ai |
What it does | What leaves your machine |
|---|---|---|
off (default) |
axe-core and magnacoda's own tests, no AI | nothing beyond the page fetch itself |
explain |
+ plain-language "why it matters / how to fix" for the top 10 findings | rule id, WCAG criterion, selector and a shortened HTML snippet, for the top 10 only |
semantic |
+ a second Claude pass over axe-blind categories (alt-text quality, reading order, link text out of context, heading structure) | the above, plus a limited extract of the page: images' alt text and file names and whether each shows, the heading outline, and link text and context, with which links go to the same place (not their addresses). Never the full page HTML or a screenshot |
Every AI-sourced finding is tagged source: "claude-semantic" (rather than "axe-core"), so you can filter the AI's judgment calls out entirely in CI, for example with jq 'select(.source=="axe-core")' on the results file.
The AI tiers are off unless you ask for them, and magnacoda asks for your consent before sending anything (in a script, -y or --yes gives it in advance). Without AI, no AI libraries are needed, and with --ai=off nothing about your scans leaves your machine. The only other connections magnacoda makes are a once-a-day check for updates and news, which sends nothing about you or your scans and can be turned off; see Updates and news.
API key
magnacoda looks for your Anthropic API key in $ANTHROPIC_API_KEY first, then falls back to a key saved via magnacoda setup (which prompts for one, hidden and never echoed, and saves it to a per-user settings file outside the project folder; see SECURITY.md (opens in a new tab) for the exact path and permissions). Remove a saved key with magnacoda setup --forget-key. If a key stops working (expired, revoked), a scan that hits that error offers to let you paste in a replacement on the spot instead of failing outright.
Two ways to run the AI tiers: an API key or your Claude subscription
--ai-engine chooses how the AI calls are made:
api: an Anthropic API key, as described above. Billed by the API per use.claude-code: your Claude Code login, so a Claude subscription (such as Pro or Max) can be used instead of a key. magnacoda runs Claude Code's official headless mode (claude -p) in the background, with no tools, no saved session and your own Claude Code hooks switched off, from an empty temporary folder, so it reads and writes nothing outside that call. You need Claude Code installed and logged in once. To log in, open a normal terminal window (not the magnacoda shell), typeclaudeto start Claude Code, then type/loginthere and follow the steps. Usage counts toward your Claude plan's limits instead of an API bill. Each call also carries about 13,000 to 16,000 tokens of Claude Code's own setup;explainmakes one call per scan andsemantictwo, times the number of scans (--screen both,--color-scheme both).auto(the default): the choice you made. When Claude Code is installed,magnacoda setupasks "Use your Claude plan for the AI tiers? [Y/n]" and saves the answer asengine = claude-planorengine = api-keyunder[ai]inpreferences.ini(run setup again, or edit that line, to change it). If you haven't chosen yet, the first AI scan asks the same question once. Your Claude plan is never used until you say yes. So without a choice, and with nobody there to answer (a script or CI), magnacoda uses an API key if there is one, and otherwise stops and says how to choose.
What differs for your data: the same bounded data is sent either way (see the table above), but with claude-code your Claude plan's terms and data settings apply, which can differ from the API's (for example, whether conversations may be used to improve Claude). The consent prompt says which applies before anything is sent.
Platforms: tested on Linux; macOS and Windows are designed for but not yet tested. Using a subscription this way is intended for your own scans; for heavy automated use such as CI, check Anthropic's current terms, or use an API key.