EGRESS LOCKED · 5 HOSTS REDIRECTS REFUSED SHELL NONE TARGET EXECUTION STRUCTURALLY IMPOSSIBLE
chainoftrust.dev · install-time trust reports INTAKE · FORM 1
Never executedstructural, not policy

Should I let my agent
install this?

Paste a repository URL. You get a report on what installing it would actually do: what the install path executes, whether published checksums are ever verified, and what gets written into your agent's configuration.

Free · no account

Specimen: surveyed by reading, never by running

Accepts a GitHub URL, owner/repo, a commit or branch URL, npm:name or pypi:name. Verdicts are public.

Plan of inquiry: six views
VIEW A-1

Install path

Does an installer exist, does the release publish checksums or signatures, and can the verification code actually run?

VIEW A-2

Agent config

Hooks, skills, MCP registrations or plugin manifests that reach an agent's harness before any install step.

VIEW A-3

Registry provenance

Does the registry entry bind the published files to the build that produced them?

VIEW B-1

Blast radius

Which environment variables and hosts the install path touches, and are any credential-shaped?

VIEW B-2

Trust root

How concentrated are publishing rights, and what is the real lineage?

VIEW B-3

Unauditable surface

How much of what you would receive cannot be reviewed by reading it?

Method of survey
no clone: files read one at a time over HTTPS no shell, no filesystem, no process spawning repository prose: nonce-fenced, treated as data fetch budget + per-file byte cap
What this is not
Not a vulnerability scanner. Where an existing tool is authoritative it is cited, not duplicated: OpenSSF Scorecard for maintenance hygiene, Socket for registry alerts. The question here is what a piece of software asks permission to do when you install it. Two of these views have no equivalent in either tool: install path reachability (does the verification code actually run) and agent config auto-discovery (what reaches an agent's harness before any install step).
Example reports, not an index

A handful of reports we picked to show the shape of one: what the annotations look like, what the chain of trust draws, what the install-path schematic looks like when there is an installer to draw. Not everything analysed and not an activity feed, just examples.

EXAMPLE

ollama/ollama

Installer runs with elevated privileges and never checks the checksums its own release publishes.

View the report →

EXAMPLE

anthropics/claude-code

Ships hooks, a plugin manifest and ten skill definitions: agent config reaches a harness before any install step.

View the report →

EXAMPLE

astral-sh/uv

A Python build backend runs at install time, alongside Claude Code hooks committed in the same repository.

View the report →

EXAMPLE

modelcontextprotocol/servers

An MCP server registration, and a package manifest with no install-time lifecycle scripts at all.

View the report →

Project
chainoftrust.dev
Drawing
Intake · Form 1
Scale
NONE. Nothing executed
Drawn by
deterministic collectors
Checked by
arithmetic, recomputable
Disputes
contact@chainoftrust.dev