- Developers
- OpenAPI
API Contract Drift Analyzer
Compare two versions of an OpenAPI 3.0, 3.1 or 3.2 spec and see which changes break existing clients: removed operations, new required fields, type, enum, response and security changes, each explained.
Turn browser network captures into evidence about APIs, performance, security and media, with every finding traced to the requests behind it, all locally in your browser.
Drop HAR files here
HAR 1.1 or 1.2 (.har), up to 200 MB per file and 10 files at once. You can open several captures and compare them.
Your HAR files are processed locally in your browser.
Background workers are not available here (for example when the page is opened from disk), so files are analysed in the page itself. Large files may make the page pause.
Every entry in the file has exactly one outcome, so the counts add up to the number of entries. Select an outcome to see the entries and the reasons.
| Domain | Party | Requests | Transferred | Failed | Median time | Longest |
|---|
| Method | URL | Type | Category | Party | MIME type |
|---|
Tokens, keys, passwords, cookies and personal data are masked. Tick “Show sensitive values” only where nobody else can see your screen.
The search and filters on the Requests tab also apply here.
Hover over a row to preview its phases; select a row to keep its details here.
An endpoint is the method, host and path of a request. Path segments that look like IDs (numbers, UUIDs, long hex or letter-and-digit tokens) are shown as {id}; other segments are kept, so endpoints are never merged just because they share a prefix. Query strings are not part of the endpoint, and URLs are masked.
| Endpoint | Category | Party | Requests | Failed | Statuses | Median | Longest | Transferred |
|---|
A value link means a value returned by an earlier response (shown by where it was, never by its content) was sent again in a later request, after that response had finished. Redirects and initiators are recorded by the browser. Requests that only happened close together in time are not linked.
Shapes list field names and value types only, never values. A GraphQL response with HTTP 200 can still contain errors, so GraphQL outcomes are read from the response body; when the body was not recorded, the outcome is shown as unknown, not as a success.
A 4xx or 5xx response is not always an application defect: sign-in flows, missing optional files and rate limits produce them too. Each group lists possible causes.
These are indicators from one captured session, not proof of a vulnerability. Values are never shown; cookie and header names are.
The original file is never changed. The copy keeps the HAR structure, with sensitive values replaced by REDACTED. Preview the changes, then download.
Field names in JSON or form bodies, separated by commas. The whole value is replaced.
Exact text by default, replaced everywhere in the file. Up to 20 patterns.
The options changed after this preview. Preview again before downloading.
Compare two captures of the same user flow, for example before and after a deployment. Only the traffic in the captures is compared: a difference shows what happened in these sessions, not why.
Open a second HAR file to compare: .
Durations and waiting times are compared as medians of responses that did not come from the cache.
HAR files often contain passwords, tokens and personal data, so they are read inside this browser tab, in a background worker that has no network access. This page’s security policy blocks background network requests, and nothing is saved: closing or reloading the page discards everything. After the page has loaded, it also works offline.
A HAR (HTTP Archive) file is a JSON record of the network requests a browser made while you used a page: every URL, method, status code, header, cookie and timing, and often the request and response bodies too. Support teams ask for HAR files because they show exactly what happened between the browser and the servers.
That completeness is also the risk. A HAR file can contain session cookies, access tokens, API keys, passwords typed into forms and personal data from API responses. Anyone who receives it may be able to use them.
Recent versions of Chrome export a sanitized HAR by default, without cookies and authorization headers. Exporting with sensitive data has to be turned on in the developer tools settings.
A HAR file uses -1 for phases that do not apply or were not measured. The tool shows these as “not recorded” and never estimates them.
No. They are read inside your browser tab, in a background worker that cannot make network requests, and the page’s security policy blocks background requests. Nothing is stored: closing or reloading the page discards the files. You can disconnect from the internet after the page has loaded and it still works.
It is much safer, but no automated tool can promise that every secret is gone. Detection is based on header and field names and on the formats of common tokens, keys and identifiers. Review the preview, remove bodies you do not need, and check the file for anything specific to you before sharing it.
The HAR file says so: -1 or a missing value means the browser did not measure that phase, for example because a connection was reused or the response came from the cache. The tool shows the gap instead of guessing.
Up to 200 MB and 100,000 requests per file, and 400 MB across all open files. Memory is the limit: a HAR file needs several times its size in memory once read. On phones and older computers, smaller files work best. Ticking “Disable cache” and leaving out large media while recording keeps captures small.
No. A HAR file records the requests the browser made, not clicks or intent, and not whether media was decoded or played. Journeys describe groups of related requests, and a media segment with status 200 only means it was delivered. Everything the tool infers says so, with a confidence level and the evidence.
GraphQL servers usually answer 200 and report failures in an errors array in the body. The tool reads that array for each operation, independently of the HTTP status. If the response body was not recorded, the outcome is shown as unknown, not as a success.
Because none of the rules matched. Classification uses only what is in the capture (body shapes, content types, the type the browser recorded, words in paths and host names) and never guesses. Open the request to see what was looked at.
By method, host and path. By default, IDs in paths, hashed file names, query parameter values and common cache-busting parameters are ignored, so /orders/123?_=171 and /orders/456?_=172 count as the same request. Each rule can be turned off.
Compare two versions of an OpenAPI 3.0, 3.1 or 3.2 spec and see which changes break existing clients: removed operations, new required fields, type, enum, response and security changes, each explained.
Paste your MCP server config and get the exact line of every mistake: JSON errors, wrong keys, bad args and env values, exposed secrets and client-specific problems. Works for Claude, VS Code, Cursor and Windsurf.
Remove GPS location, device details, edit history, and AI generation details such as prompts and Content Credentials from JPEG, PNG, WebP, MP4, MOV, WebM and MKV files. No quality loss.
AICEA designs and builds APIs, integrations and the monitoring and tests that keep them fast and reliable.