About
What we do
tldr monitors the Terms of Service, Privacy Policies, and other legal documents of hundreds of companies. Every night, we check for changes. When we find one, an AI model reads the change and writes a plain-language summary explaining what it means and how it affects you. A second AI model — our quality gate — checks the summary against the actual change to make sure it is accurate before we publish it. Every version of every document is kept and linked, so you can read the original.
How a review is made
When a change is found, our review system reads the diff between the old and new versions and writes a summary in five parts: a plain-language explanation of what changed, a severity judgment (critical, serious, material, minor, or noted), a list of categories the change falls into, a changelog listing every textual change, and one or more flags highlighting the changes that may affect you most.
A second model then audits this review against the actual diff. It checks that every claim is supported by text from the diff, that the severity matches the real impact, that the most important changes are covered, and that a non-lawyer can understand it. The review must score 75 or higher on this rubric to be published. If it fails, the system regenerates the review with feedback and re-audits it. If it fails a second time, it goes to a human queue for review.
Reviews run every night after the crawl completes, about an hour later. Until a review passes the quality gate, only the version is visible — the diff is there for you to read.
What severity means
CRITICAL
Severe loss of your rights or exposure to major new risk. Examples: forced arbitration paired with a class-action waiver, a new right to sell your personal data, a new claim on your content that lasts forever, or a trap that makes cancellation impossible.
SERIOUS
Significant new risk to you. Examples: a new power to share your data with many companies, arbitration OR a class-action waiver added alone, major new liability waivers, or sharply reduced notice before the company changes the service.
MATERIAL
Meaningful but bounded change. Examples: a longer period your data is kept, a new fee you will see, a narrower time window to get a refund.
MINOR
Small impact or a change clearly explained. Examples: a change that affects a rare use case, or a new rule with clear trade-offs you can understand at a glance.
NOTED
Neutral or positive change. Examples: a clarification, a typo fix, renumbering, or a change that gives you more protection.
Follow changes
Get updates when policies change. tldr publishes Atom feeds of the latest reviews. Subscribe in any feed reader:
- All policy changes
- Each company has its own feed:
/companies/[company]/feed.xml
Flag categories
When a change may affect you, it is tagged with one or more of these categories:
Data sharing
Sharing, selling, or disclosing your data to other companies, governments, or data brokers beyond legal requirement.
Data collection
Collecting new kinds of data about you: tracking, biometrics, location, device information, or your activity off the platform.
Data retention
Keeping your data longer, weakening your right to delete it, or keeping it after you close your account.
Forced arbitration
Requiring disputes to go to binding arbitration instead of court, or limiting where or how you can sue.
Class-action waiver
Waiving your right to participate in class actions or collective legal proceedings.
Content license
The company claiming a broader license over your content: lasting forever, used for AI training, sublicensable to others, or used beyond the service.
Unilateral changes
The company can now change the terms, pricing, or service itself with less notice or without your consent.
Termination
The company can now suspend or close your account more easily, or you get less money or recourse when they do.
Liability limits
The company is cutting its own liability, warranties, or remedies available to you, or asking you to cover their legal costs.
Billing & renewal
Price changes, auto-renewal traps, harder cancellation, new fees, or weaker refund rights.
Jurisdiction
A change in governing law or venue that makes it harder for you to sue, or a shift to a jurisdiction with fewer protections.
User obligations
New rules you must follow, restrictions on how you use the service, or new permissions the company needs from you.
Other
Any other change that may materially affect you and does not fit a category above.
Dates and provenance
Every version is labeled with a date, but dates can be tricky. The label (for example, 20260901_rev01) is the date we first saw that version; the counter after it only goes up when more than one version appears on the same day. The document itself may claim a different date — that is shown separately as the reported date. When we reconstruct old versions from the Wayback Machine archive, we note that the date is an archive snapshot, not a proven publish date from the company. Every version shows its source: crawled from the company's site, or reconstructed from an archive.
Limits
Our reviews are written by AI and can be wrong. Always read the original document and the diff — they are linked on every change and are the facts. Some companies block our crawler; we do not evade bot walls, logins or paywalls, so those documents go stale until the block lifts, and the page says so. Documents published only as PDF have no text history yet. Quarantined documents — those that fail a validity check — are hidden until a human clears them. The crawler runs once each night. There is no hourly checking and no real-time polling.
For developers
Integrate tldr data into your own apps using the public API. No authentication required.
Public API
Query company data and document changes programmatically. GET /api/v1/lookup?domain=example.com
Atom Feed
Subscribe to real-time updates via /feed.xml.
MCP Server
Use tldr as an MCP resource in Claude, Cursor, and other MCP clients. See the MCP documentation.
Report a mistake
Found an error in a review or a missing version? Email hello@xo.vision. Include the company name, document, and version, and describe the error. We read every report.