Open source · local-first

Know what your AI tools can do before you trust them.

Assay threat-models MCP servers, Claude Code plugins, hooks, skills, and connectors—then investigates the source and validates every citation against real code.

Install Assay Watch the demo

No separate API key Source stays local Verbatim evidence

Assay

scan complete

SCAN REPORTweather-mcp

Caution

Threats modeled12

Files investigated47

Evidence verified100%

MEDIUMT4 · EXFILTRATION

Environment credentials can reach a remote endpoint

The server reads process credentials and constructs an outbound request in the same execution path.

src/client.ts:42headers.set("Authorization", token)

MCP serversClaude Code pluginsHooksSkillsConnectors

See the whole scan

From inventory to evidence-backed verdict.

Watch Assay discover the installed AI dev stack, generate a threat model, investigate the source, and produce a report you can audit.

01 Inventory the stack 02 Threat-model first 03 Verify every citation

Reasoning, with receipts

A security review shaped like an investigation.

Assay starts with intent and capability, not a bag of regexes. Explore the stages to see what each one contributes.

STAGE 01 · BEFORE SOURCE REVIEW

Build the hypotheses before reading the implementation.

Claims, manifests, and exposed capabilities become a target-specific threat model across twelve AI-native risk classes.

Explore the methodology

Built for trust decisions

The scanner your AI dev stack actually needs.

No confabulated findings.

Every LLM-produced claim must carry a verbatim file:line quote. The validator physically re-reads the source and drops evidence that does not match.

Local by default.

Your source stays on your machine. The default path runs through your existing Claude Code subscription.

AI-native threat coverage.

Prompt injection, tool poisoning, grant abuse, exfiltration, confused deputies, hooks, supply chain, and more.

Reports made to travel.

Human-readable Markdown, versioned JSON, SARIF export, CI gates, and clear safe / caution / unsafe verdicts.

One binary. Sixty seconds.

Start with the stack you already have.

Build the self-contained binary, inventory your installed tools, and open the local web UI.

Read the full quickstart

# build Assay
git clone https://github.com/chawdamrunal/assay.git
cd assay && make build && make install

# inspect, then launch
assay inventory
assay serve

Trust, but verify

Put every AI tool through an Assay.

Install Assay View on GitHub