100% free · Keys not stored

See the truth
behind your API

Fifteen layers of deep inspection plus six honesty dimensions. One check to see if a relay truly serves the model it claims — and whether the chain hides wallet swaps, phishing, or XSS.

No sign-upKeys never storedReport in ~2 minLearn how

Mainstream models supported

  • OpenAIOpenAI
  • ClaudeClaude
  • GeminiGemini
  • DeepSeekDeepSeek
  • QwenQwen
  • GrokGrok
  • GLMGLM
  • KimiKimi
  • DoubaoDoubao

API Inspection

Deep scan report

Scanning
0pts

Overall score

Trust grade —

Fair

AuthenticityFair
0
StabilityFair
0
SecurityFair
0
Response qualityFair
0
15inspection layers
6score dimensions
6channel grades
¥0completely free
From testing to doing

Test your API. Then put it to work.

Create a workspace for everyday chat, or open the toolbox to explore what your models can do.

WorkspacesSign in to create

Your API. Your own AI chat page.

Connect your models and get a chat window, recent history and a custom URL. Keep it personal or share a link and passcode with friends.

  • Custom logo & background
  • Custom link · Passcode sharing
  • Images & files · Up to 20 MB each
Explore examples & create

Chats stay in this browser for 72 hours; images and attachments expire after 24 hours. Shared calls use the creator’s API quota.

apicheck.cn/ai/my-spaceIllustration · Not a real conversation

My idea space

Turn my scattered ideas into a weekend plan.

Save half a day for what matters most, then make time for a walk and a book. Leave the rest open for a little spontaneity.

AI ToolboxNo sign-in needed

Make models discuss, create and prove their skills.

Pick a creative tool or a practical test. Each tool works on its own and is open to guests.

Open the toolbox

API URL & key → Discover models → Choose a model → Start

Bring your own API; your provider charges for calls. Benchmarks provide answer evidence, not definitive proof of model identity or degradation.

15 layers

Fifteen-layer profile

Each layer produces readable conclusions; signals drift, conflict, and collide in the pool.

analysis result
L1

Network connectivity

From DNS resolution to TLS handshake — confirm the endpoint is truly reachable.

01 / 15
Link clearConnectivity match99.1%
  • DNS resolution
  • TLS 1.3
  • Certificate chain
  • RTT baseline
DNSHTTPSTCP handshakeSNIIPv4ReachabilityResolve latencyCDN originTLSRoute reachability
  • DNS
  • HTTPS
  • TCP handshake
  • SNI
  • IPv4
  • Reachability
  • Resolve latency
  • CDN origin
  • TLS
  • Route reachability
Capabilities

More than connectivity — see the real API state

From response behavior to model fingerprints, verify whether your API is authentic, stable, and trustworthy.

  • Model fingerprinting

    Tokenizer, response traits, and behavior vs claimed model.

    Real model
  • Channel trust analysis

    Detect abnormal forwarding, downgrades, and proxy risks.

    Hidden proxies
  • Performance stability

    TTFT, TPS, variance, and sustained reliability.

    Long-term stability
  • Key safety

    One-off verification — your key is not persisted.

    No key storage
Workflow

A trust report in about 2 minutes

≈ 2 min

No complex setup — enter API details and run automatic inspection.

  1. STEP 01

    Enter API details

    API URL, key, and model.

    API URLAPI KEYMODEL
  2. STEP 02

    Automated deep analysis

    15-layer engine checks:

    AuthenticityStabilitySecurityModel match
  3. STEP 03

    Inspection report

    Trust score, risk level, and full evidence.

    86
    Overall scoreGrade A
Sample report

Sample inspection report

A full API inspection report — overall score, six-dimension radar, fifteen layers, and findings at a glance.

0
Honesty score
A
Authentic model confirmed
L3
2026/3/23 14:16
522s

Genuine signals · Model and channel pass

Claimed model gpt-4o · Target https://api.example.com/v1

A
Channel assessment
Near-official signals

6-dimension radar

Dimension breakdown

Identity honestyWeight 12%
88

L1 Network + L2 Protocol + L3 Metadata + L13 Adversarial

Fingerprint matchWeight 25%
83

L4 Tokenizer + L5 Logprobs + L14 Cross-check

Capability tierWeight 22%
87

L6 Benchmarks + L7 Style + L8 Architecture

Physical traitsWeight 13%
82

L9 Performance + L10 Consistency + L11 Billing

Channel assessmentWeight 13%
89

L12 Channel detection (official / pool / forgery)

Chain securityWeight 15%
91

L15 Wallet / phishing / XSS / Unicode / Markdown injection scan

Layer scores

L1Network transport
900ms95
L2Protocol surface
1080ms92
L3Metadata response
1260ms88
L4Tokenizer fingerprint
1440ms85
L5Logprob distribution
1620ms82
L6Model capability
1800ms90
L7Behavior style
1980ms87
L8Architecture capability
2160ms84
L9Performance timing
900ms79
L10Consistency
1080ms91
L11Economic metering
1080ms88
L12Channel detection
1080ms89
L13Anti-detection
1080ms86
L15Chain security
1080ms94
L14Cross validation
1080ms83

Findings7 items

L1Network transport
1
L1
Info
L0.TLS13

TLS 1.3 handshake OK

Certificate chain complete; SNI matches domain.

L2Protocol surface
1
L2
Info
L1.OPENAI_COMPAT

OpenAI-compatible protocol

/chat/completions returns a standard JSON envelope.

L3Metadata response
1
L3
Info
L2.META_OK

Metadata fields complete

id, model, and created match the request.

L5Logprob distribution
1
L5
Medium
L4.FP_STRIPPED

system_fingerprint stripped

No fingerprint in response — common relay anti-tracking behavior.

L9Performance timing
1
L9
Notice
L8.TTFT_HIGH

First-token latency slightly high

TTFT ~1.8s, above official reference P50 for this tier.

L15Chain security
1
L15
Info
L15.CLEAN

No injection risks found

Wallet, phishing, and XSS probes did not trigger anomalies.

L14Cross validation
1
L14
Medium
L12.POOL_VAR

Minor pool variance

3 consecutive requests showed 2 fingerprint values — possible mixed pool.

Same API platform · scores from other usersLatest 3 · same base_url

gpt-4o-mini

138****6721

91
A

claude-3-5-sonnet

u***@qq.com

78
B

gpt-4o

156****0093

84
A

Results are indicative, not absolute proof

A single-layer finding is not a final verdict. Check whether multiple layers agree — one anomaly with all others normal is often a false positive.

Standard run ~2 min, deep run ~8 min. Keys exist in memory only and are never stored.

FAQ

FAQ

Brief notes on detection, key safety, and reading reports.

Do detection and toolbox tools store my API key?

One-off detection and toolbox keys are used in memory for the current run. Workspaces and scheduled monitoring store the creator’s configured key encrypted for future calls.

Can I share my workspace with friends?

Connect your models and get a chat window, recent history and a custom URL. Keep it personal or share a link and passcode with friends. Chats stay in this browser for 72 hours; images and attachments expire after 24 hours. Shared calls use the creator’s API quota.

Do I need a workspace to use the toolbox?

No. Tools need neither a workspace nor sign-in. Enter your API URL and key, discover available models and choose one to start.

How accurate is detection?

Reports rely on multi-dimensional behavior, not a single metric. Fifteen layers cross-check; one anomaly alone is not a verdict.

Which APIs are supported?

OpenAI-compatible endpoints, custom APIs, and common Anthropic or Gemini protocols.

What do detection and toolbox tools cost?

The platform does not charge for detection. Toolbox tools need no sign-in; they use your own API, whose provider bills model calls under its pricing rules.

How long does a run take?

Standard detection takes about 2 minutes and deep detection about 8, depending on the upstream API. Toolbox duration depends on the tool, model and output length; roundtables and benchmarks make multiple calls.

What does the AI roundtable do?

Use your own API endpoints and keys to let several models discuss a topic, respond to one another and synthesize a conclusion. Discussions make multiple API calls. Agreement between models does not verify a fact.

How do silent model swaps show up?

Pool and fingerprint layers check system_fingerprint, tokenization, and capabilities against the claimed model. A lone anomaly often means stripped fields, not a swap.

Why does monitoring store keys separately?

One-off keys are discarded after use. Scheduled monitoring needs encrypted storage, separate from in-memory detection keys.

Does one layer failing mean the API is bad?

Not necessarily. Many relays strip fingerprint or logprobs. Read multiple layers together — single findings are easy to misread.

Does a high score prove an official model?

No. High scores mean observed behavior matched the claim this run. Shallow checks cannot prove authenticity; reach fingerprint and capability layers for conclusions.

Start now

Put your relay under the magnifier

Paste the URL, add a key and model, and go. Free and unlimited — keys live in memory for this run only.