tracingviolet
Book an audit

Are your endpoints agent-ready?

Can agents actually use your MCP server or API? We test with real models, show where they fail, and tell you exactly what to change.

See how it works

This is how we see your tools.

We don't lint your schemas or score your docs. We watch real agents do real work, and trace what your logs can't see.

session-04d5 · arvossi.docs · claude-opus-4.8
illustrative session
00:00.01[sys ]trace.start → target: api.arvossi.dev/docs · model: claude-opus-4.8
00:00.13[invoke ]agent selected tool docs.create
00:00.27[present]tool schema loaded · 6 params · description 144 tokens
00:01.04[friction]input formed · 0 retries · clean
00:02.88[exec ]POST /v1/docs → 200 · 412ms
00:02.94[utility]result parsed · prose response · no artifact_id
00:05.10[agent ]▲Done. Let me know if you need anything else. — completion verifiable? false
00:05.11[fix ]
recommended fix → wrap result in { artifact, status }
00:05.14[ready ]awaiting next target…
stage hit:  outcome · task completion verdict:  incomplete recommended fix:  wrap result in { artifact, status } loop · 01 / 04

the audit traces hundreds of live sessions across your service — and reads them against thousands we've traced across other services.

// backed by independent research

Even the best AI model fails nearly one in five tasks when using production MCP servers.

MCP-Atlas, Scale AI — Bandi et al., 2026 · arXiv 2602.00933 · 36 production MCP servers, 20 frontier models

97.1% of tools have at least one description quality issue.

The description layer is what agents read before they ever call you.

Hasan et al., 2026 · arXiv 2602.14878 · 103 MCP servers, 856 tools

Standard-compliant tool descriptions reach 72% selection probability vs a 20% baseline.

Wang et al., 2026 · arXiv 2602.18914 · 10,831 MCP servers

89% of developers use generative AI daily, yet only 24% design APIs for AI agents.

Postman, 2025 State of the API Report · 5,700+ developers, architects, and executives

// tracing the agent's journey

Losing agent traffic? Find the break. Ship the fix.

// the five diagnostic stages
01
Invocation
Did the agent try to use the tool at all?
02
Presentability
Could the model load your tools without hitting platform limits?
03
Input Friction
Could the agent reach a useful call without prep work?
04
Reliability
Did the call return successfully?
05
Utility
When the call succeeded, was the response useful?
→
// the outcome

Task Completion

Did the user get what they came for?

Your logs can't see this

We don't just diagnose.
We tell you exactly what to change.

F-01 FIX Return a structured result from docs.create API change illustrative

Replace the prose confirmation with a structured object the agent can verify and cite.

Current
"Created your document! View it at arvossi.dev/d/8f3k2"
Revised
{ "artifact": { "id": "doc_8f3k2", "url": "https://arvossi.dev/d/8f3k2" }, "status": "created"}
Done when
docs.create responses carry artifact.id and status — and the agent's final answer can cite them.

Every finding maps to a specific fix — exact schema, exact text, exact error format. Precedent from patterns we've measured, not guesswork.

Ready to find out what's breaking?

We test your tools live across multiple models and deliver a report with specific fixes you can ship.

or hello@tracingviolet.dev

Not ready for a full audit?

Run our free Agent Readiness Scan. See what your tool surface looks like from the agent's perspective.