ElkwaiAPI change review
Flag removed operations, newly required parameters or bodies, removed response codes and changed security declarations in OpenAPI 3.0/3.1 JSON.
BRING YOUR OWN AI
Developer Assurance Kit
Your AI. Your interfaces. Checks you can repeat.
Keep using your own model, agent and infrastructure. Compare MCP and API contracts, test recorded AI outputs against explicit expectations, and retain source-hashed reports. No hosted AI plan or custom integration engagement required.
Five local check modes. Free browser: five payload or regression cases. No account, API key or upload. No per-run charge or model inference from Elkwai Solutions.
01 Five checks, one shared engine
Payload validation
Check example payloads against a documented JSON Schema 2020-12 subset, without coercion, inserted defaults or remote fetching.
Schema change reports
Inspect added, removed and changed schema fields with JSON-pointer locations, exact input hashes and rule versions.
MCP contract changes
Compare complete tool catalogs by name. Catch removed tools, newly required arguments, changed output contracts, descriptions and behavioral hints. Hints are not permissions.
Try MCP contract reviewAI output regression
Check observed runs against exact values, schemas, required source quotes and entered cost/latency budgets. Missing runs and measurements fail instead of disappearing from the score.
Try recorded-run checks02 Carry the review into your build
| Surface | Working output | Boundary |
|---|---|---|
| Offline browser | Editable JSON, severity filtering, source receipts and JSON export | Session-only inputs; no automatic save |
| Node.js CLI | JSON to stdout, high/review exit thresholds and isolated workers | Node.js 22+; no hosted runner included |
| JavaScript SDK | The same versioned engine and report schema | Caller supplies isolation for untrusted input |
| CI example | Private-runner integration and explicit exit codes | Customer owns CI configuration and costs |
node cli.mjs api before.json after.json --fail-on high
Exit 0: the chosen threshold was not reached. Exit 1: findings reached the threshold. Exit 2: input or runtime failure; no passing report. Unchanged definitions do not fail merely because the tool discloses its limits.
03 Own the version you use
Developer Assurance Kit 1.1.0
A signed-release ZIP with the offline browser, bundled CLI and SDK, local MCP server, CI example, illustrative fixtures, methods, license and third-party notices. No package installation required to run the included tools.
- Up to 100 payload or AI regression cases; free workspace supports five
- Five read-only local MCP tools for your compatible AI client
- Same API, schema, MCP and regression rules across all surfaces
- Source SHA-256, versioned JSON reports and CI exit thresholds
- Five-second isolated checks; one concurrent MCP check
- One organization's internal development and authorized client analysis
$79one timeIntroductory pricing
Purchased version only. Private CI and local MCP included. No subscription, per-run charge, hosting or promised future versions.
Checking purchase and private delivery availability.
Tax at checkout. Retrieve the ZIP within 30 days; keep using the downloaded version offline. Review the license, purchase terms and privacy notice.
04 Connect checks, not another AI provider
For agent and MCP builders
Capture the complete tool catalog before and after a release. Review contract changes, then replay your own representative consumer cases. The Kit never invokes the tools being inspected.
For document and RAG workflows
Retain expected fields and source excerpts. Compare recorded outputs and quoted evidence when prompts or models change. Exact quote membership is not a judgment of truth or entailment.
For cost-aware teams
Set per-case latency in milliseconds and cost in millionths of USD. Check your recorded measurements against those limits. The Kit does not call providers or measure bills independently.
For your existing AI client
Launch the included stdio MCP server with Node.js 22+. Five schema-declared tools return structured findings. No server port, API key, shell tool or file-reading tool is exposed. Your AI client's data handling still applies.
node cli.mjs regression expected.json observed.json --fail-on high
node /absolute/path/to/extracted-kit/mcp-server.mjs
05 Specific methods and practical limits
Bounded JSON processing
512 KiB/input, 32 nesting levels, 20,000 JSON nodes, 512 nodes per schema, 500 API operations or MCP tools and 1,000 findings. Browser/CLI/MCP checks stop after five seconds. Duplicate keys, unsafe integer values and unsupported schemas are rejected, not silently accepted.
No source uploads
Browser inputs stay in memory. No API calls, reference fetching, telemetry or autosave. Reports omit raw input but can contain confidential property names and paths. Review exported reports before sharing. Your AI client sees the content it submits to local MCP tools.
Documented schema subset
Types, bounds, object and array constraints, compositions, conditional schemas and acyclic local references are supported. YAML, formats, content decoding, dynamic or recursive references and unknown keywords are not included.
Not a security certification
No full OpenAPI/MCP conformance or consumer compatibility proof. Exact regression expectations do not prove semantic accuracy or freedom from hallucinations. SDK consumers must provide their own worker and resource isolation.