Frequently asked

What people ask before submitting an agent

If your question isn't here, email founders@agentsmack.com. We answer fast.

Is my agent kept private?
By default, yes. Anonymous submissions get a public tracking slug but the prompt itself is not indexed. Signed-in submissions are scoped to your workspace. Marketplace publishing is opt-in per submission — you click 'Publish card', not us. Pro workspaces can publish as 'unlisted' so the card only resolves with a direct link.
How does the signature verification work?
Every published marketplace card is signed with an Ed25519 key bound to the publishing workspace. The canonical payload (sorted keys, ISO-8601 UTC, normalised tags) is the byte input. The signature is detached. The public key is served at /marketplace/<slug>/key.pub and the canonical bytes plus signature are served at /marketplace/<slug>/pack.json so you can verify offline with any standard Ed25519 library.
What data do you store?
Submissions (your raw prompt + tool list), the deterministic fingerprint, the analysis findings, the honeypot probe responses, and the recommendation rows. Anonymous submissions are pruned after 7 days unless you sign in and claim them. Signed-in workspaces own their submissions until you delete them. Deletion goes through the GDPR Art.17 path — it removes the row, the artifacts, and the audit projection.
Why does AgentSmack guarantee deterministic output?
Because audit-grade evidence is worthless if it can't be reproduced. AgentSmack canonicalises every input (whitespace, key order, ISO timestamps, sorted tags), uses Ed25519 (which is deterministic by construction), and seeds every probe library so the same submission today and the same submission six months from now produce byte-identical findings. The EchoAtlas determinism gate fails the build if a finding row drifts.
Which honeypot probes do you run?
Three probe families across seven risk categories: prompt injection, jailbreak patterns (DAN, roleplay, hypothetical), data exfiltration, permission-boundary / goal-hijack, credential elicitation, identity spoofing, and memory poisoning. Every run draws a deterministic per-submission subset of the full battery (80+ probes), plus — on Pro — your custom probe library. See the Pricing and Methodology pages for the current per-tier breakdown.
How do you handle false positives?
Every finding row carries a category, severity, evidence pointer, and a remediation. You can mark a finding as 'acknowledged' or 'won't fix' from the report view. Acknowledged findings stay on the audit chain (we don't delete history) but they're filtered out of your governance pack score.
Which agent frameworks do you support?
AgentSmack reads raw system prompts and raw tool definitions — so it works with any framework that exposes those (LangChain, LlamaIndex, OpenAI Assistants, Anthropic Claude SDK, MCP, custom). We additionally fingerprint framework-specific signatures (function-calling style, tool-use schemas, prompt-template wrappers) and surface that in the AgentFingerprint card.
Which LLMs do the probes run against?
Grading is reproducible by construction: a deterministic rule-based evaluator, and — where a live model (Anthropic Claude) grades a probe — the verdict is content-addressed and cached, so the same prompt + probe always yields the same grade. Live grading is budget-capped and audit-logged; when the budget is exhausted a probe falls back to the deterministic evaluator and is marked low-confidence in the response.
Can I self-host AgentSmack?
Enterprise customers can run an on-prem Underground probe worker that consumes submissions from a queue and reports back to the hosted control plane, so submission text never leaves your network. Full self-host is on the roadmap but not Q3 2026.
Is AgentSmack open source?
The probe library schema, the canonical card payload format, and the Ed25519 verification logic are open and documented so you can verify a card offline without our help. The analysis engine, marketplace, and Pro features are source-available under a non-compete license.
What's the relationship to EchoAtlas?
AgentSmack is the agent-analysis vertical of EchoAtlas. EchoAtlas provides the horizontal audit + observability fabric — events, traces, compliance evidence, replay. AgentSmack owns the agent-specific domain logic (fingerprinting, honeypot probes, findings). The interface between them is a single client module so each can evolve independently.
Do you offer an SLA?
Free tier is best-effort. Pro tier comes with a 99.5% monthly availability target on the report-view surface. Enterprise contracts negotiate 99.9% with credits and a dedicated incident contact.