Agent package risk radar.

Review required for agent use. This is a metadata and execution-boundary card, not an accusation.

MCP metadata is public, but tool scope and credential exposure still need a downstream trust decision.

Verdict review, risk 88/100, attention 96/100.

Decision path: /api/decision?q=MCP%20server%20with%20filesystem%20and%20token%20tools&sources=mcp&limit=5

Review required before install. The card describes an execution surface, not a malware verdict.

Install-time hooks are one of the easiest ways for a package to execute before an agent has reviewed the code.

Verdict review, risk 86/100, attention 94/100.

Decision path: /api/decision?q=npm%20package%20with%20preinstall%20script&sources=npm&limit=5

Review the identity binding. Similar names and copied metadata require human or host confirmation.

Typosquatting works because agents often trust names, versions and copied repository URLs too early.

Verdict review, risk 82/100, attention 91/100.

Decision path: /api/decision?q=OpenSearch%20JavaScript%20client%20npm%20package&sources=npm&limit=5

Review package provenance before agent execution. Popularity is not execution permission.

Agent installs need provenance and maintainer continuity, not popularity alone.

Verdict review, risk 68/100, attention 81/100.

Decision path: /api/decision?q=Python%20HTTP%20client%20package&sources=pypi&limit=5

Review required for model loading. This is not a model quality score or safety guarantee.

Model selection is now package selection: loaders, custom code and weights need the same preflight receipt.

Verdict review, risk 74/100, attention 84/100.

Decision path: /api/decision?q=Hugging%20Face%20model%20requiring%20remote%20code&sources=huggingface-model&limit=5

Review required before automated dataset loading. The radar does not judge dataset content quality.

Datasets are not passive files when the loading path executes scripts or fetches additional resources.

Verdict review, risk 63/100, attention 78/100.

Decision path: /api/decision?q=Hugging%20Face%20dataset%20with%20loading%20script&sources=huggingface-dataset&limit=5

Review repository instructions before execution. Nipmod does not post unsolicited issues or comments.

Repositories increasingly include agent-facing instructions, but README text is not a host policy.

Verdict review, risk 71/100, attention 83/100.

Decision path: /api/decision?q=GitHub%20repository%20with%20install%20instructions%20for%20agents&sources=github&limit=5

Review required before runtime. A pullable image is not permission to run it.

Agent tool recommendations often cross from package metadata into container runtime permissions.

Verdict review, risk 69/100, attention 76/100.

Decision path: /api/decision?q=Docker%20image%20for%20agent%20tool%20runtime&sources=dockerhub&limit=5

Review extension permissions before enabling. The card is a preflight checklist, not an endorsement.

Agent tooling often enters through editor extensions, where installation creates a long-lived local execution surface.

Verdict review, risk 66/100, attention 72/100.

Decision path: /api/decision?q=VS%20Code%20extension%20for%20AI%20coding%20workflow&sources=openvsx&limit=5