MCP Registry

@everworker/oneringai

70f48c4e-e6e0-5b6c-bdcd-56fde6ac7df5 | Store v1.2.1
100/ 100
CRITICAL risk
No change since v1.2.0
AI review verdict
Do not install

The AI review rates the findings as likely false positive, but the risk score (100/100) still counts them.

Analysis record

Analysed
3 weeks ago 29 October 2026
Version
v1.2.1
Artifact
SHA256 b524d8a2…c3abc2
Detail
Full list

Is @everworker/oneringai safe?

@everworker/oneringai is a TypeScript library for building AI agents: it wraps several model providers, handles tools, memory and media, and plugs into agent runtimes. The package declares no permissions and no host permissions, and no particular web destination was recorded against it. Its network code lives in the compiled bundle at dist/index.js, where the scan found fetch calls at lines 587 through 34720, which fits a library built to call model provider APIs.

The loudest findings are eight critical MCP-TRANSPORT-HARDCODED-TOKEN hits in dist/index.js, dist/index.cjs, dist/capabilities/images/index.js and dist/agent-runtime/index.js. If one of those were a real embedded secret, it would mean the publisher shipped an access key inside the build. The same rule firing in four separate modules, in both the .js and .cjs version of each, reads like a match on the code that sets up MCP transports and auth headers rather than one key copied around. Nothing in the scan touches a credential file: no .ssh keys, no .aws/credentials, no .kube/config.

That second point decides it. A package that steals credentials has to read them and then send them somewhere. Here the scanning found no read of those paths and no unknown destination to send anything to, and the fetch calls sit in a bundle whose purpose is talking to model providers.

The scanner also flagged fetch in dist/index.d.ts, a type declaration file that contains no running code. Matches there come from type signatures, which tells you the network rule is matching the word fetch rather than a live request.

Evidence ledger

Ranked by severity · findings with a source location link to the code viewer

105 findings
Showing 25 of 105 · highest severity first

Finding Categories

8
Secrets
90
Network

MCP Server Analysis

MCP servers expose tools and resources to AI assistants. Unlike browser extensions, they run as standalone processes with direct system access. Tool definitions are analyzed for prompt injection, data exfiltration, and tool poisoning patterns.

AI Security Report

AI Security Review

Evidence context: threat category none; evidence quality moderate.

@everworker/oneringai is a TypeScript library for building AI agents: model wrappers across several vendors, tool calling, memory, media handling and pluggable OneRingAI/Codex agent runtimes. Only compiled output was available, so every finding here points at build artifacts under dist/.

Tool poisoning is the first thing to check with anything MCP-shaped, and the scan produced no matches. That zero carries weight, because this library does contain tool registration code with name and description fields for the tools it hands to a model. Rules aimed at tool metadata would have lit those up. Nothing did, so there are no hidden directives aimed at an agent, no invisible Unicode payloads and no "silently run this" phrasing in tool descriptions.

The credential findings deserve the attention they got. Eight critical MCP-TRANSPORT-HARDCODED-TOKEN hits land in dist/index.js, dist/index.cjs, dist/capabilities/images/index.js and dist/agent-runtime/index.js, with the .cjs versions of each alongside. The same rule firing in four separate modules across both module formats is the shape of a pattern match on transport setup code. An SDK that implements stdio and HTTP transports has to build auth headers somewhere, and the rule looks for token-shaped strings near that code. None of the findings touch a credential path: no .ssh, no .aws/credentials, no .kube/config, no application_default_credentials.json. The package declares no permissions and the scan recorded no destination domains, so there is no harvest side and no exfiltration side to pair up.

The 90 NET-FETCH findings sit in dist/index.js at lines 587 through 34720 and, more tellingly, in dist/index.d.ts. A declaration file holds types and cannot execute a request, so the fetch matches there come from identifiers such as a fetch property in a client interface. The runtime hits line up with a library whose entire job is calling model provider APIs.

The strongest case against this conclusion: a single-maintainer package, no linked repository, outbound calls, and a critical-severity secret rule firing eight times. If one of those matches turned out to be a live API key literal, it would be the publisher's own key in their own build, which is a hygiene problem for them. It would not be a route to a user's AWS keys, because nothing in the scan reads those paths. Exfiltration needs both halves, and only the harmless half is present.

Key Reasons

  • No tool-poisoning matches despite the library containing tool registration code, and no credential-path findings (.ssh, .aws, .kube) anywhere in the scan
  • Eight MCP-TRANSPORT-HARDCODED-TOKEN hits repeat across dist/index.js, dist/index.cjs, dist/capabilities/images/index.js and dist/agent-runtime/index.js, the pattern of a rule matching transport auth setup in both module formats
  • NET-FETCH matches in dist/index.d.ts show the network rule firing on type declarations that cannot execute requests
  • The 90 network findings sit inside bundled dist/ build output for a library whose stated function is calling model provider APIs, and no destination domains were recorded
  • No harvest-plus-exfiltration pairing: credential reads are absent and network calls go to model providers, not unknown infrastructure

False Positive Considerations

  • NET-FETCH matches against the fetch identifier in dist/index.d.ts type declarations
  • MCP-TRANSPORT-HARDCODED-TOKEN matching MCP transport header construction code duplicated across .js and .cjs builds
  • High network finding volume from bundled provider clients in dist/index.js build output

Reviewed 2026-10-31 · version 1.2.1; recommended action: suppress false positive; model confidence 73%.

MCP version history

Risk trend by version

3 analyzed versions. Each point is the latest successful scan for that version; failed zero-score scans are hidden. Dates: First seen by Risky Plugins.

Selected
100
Change since first
+35
Change from previous
No change
Versions:
First analyzed version
0.5.0
First seen Mar 17, 2026
Risk range
64 to 100
Across analyzed versions
Latest analyzed version
1.2.1
First seen Oct 29, 2026
Selected version
critical
Version
v1.2.1
First seen 3 weeks ago
Risk score
100
Findings
105
Change vs previous
No change

Pick any point on the chart, then open the Source tab to read that version's code.

Source Code Viewer

Source Code Not Available

Source code is not available for this version of the extension.

About This Extension

Unified TypeScript AI agent library with multi-vendor models, tools, memory, media, and pluggable OneRingAI/Codex agent runtimes

Frequently Asked Questions