modelwatch-mcp

MCPattested
v0.1.3io.github.bch1212MITUpdated 5d agonpmGitHub

MCP server for ModelWatch — continuous behavioral drift monitoring for LLM-powered applications.

Works in
ClaudeCursorCopilotGemini

Inferred from the transports this listing declares (stdio). A client not listed here hasn’t been ruled out — it just isn’t something Forge can confirm.

Attested build
A verified provenance attestation binds this artifact to the listed repository. Nobody has claimed the listing yet — this proves where the code was built, not who stands behind it.
186Downloads/wk
5d agoLast update
Needs 1 credential before it runs
  • MODELWATCH_API_KEYAPI keyrequired

    API key from https://modelwatch.app (free tier: 5 specs, 500 runs/mo).

Declared by the author in the official MCP registry. Forge does not store, broker, or ever see these values — the config below is scaffolded with placeholders you fill in locally.

Package
Authorio.github.bch1212
LicenseMIT
Version0.1.3
Sourcenpm+mcp-registry
npmmodelwatch-mcp
GitHubbch1212/modelwatch
SBOMCycloneDX 1.6 ↓
Trust Status
A
85/100Trusted
✓Listed in Forge index+10/10
✓Identity verified · attested build+20/20
—Ed25519 publish signature+0/5
→ Included automatically when the publisher runs `forge publish`
—Domain verification+0/5
→ Publisher: host /.well-known/forge.json on the package homepage with { "publisher": "<github-login>" }
✓npm Trusted Publishing (Sigstore)+5/5
—npm maintainer match+0/5
→ Publisher: add the verified GitHub login to the npm package's maintainers (npm owner add <login>)
✓CVE scan · clean+30/30
✓Static analysis · clean+20/20
Paste into Claude Code, Cursor, or any AI assistant to fix all gaps
StatusIdentity verified
PublisherUnverified
SignatureUnsigned
Domain—
Provenance✓ Sigstore-verified · 76d6acb
Dependencies✓ 60 resolved+ · none vulnerable
Tool surface9 tools · none privileged
Security scan✓ Cleanv0.1.3 · todayHow well does this scan work?
EvalsNone
IndexedOct 10, 2026

Verification confirms publisher identity (repo ownership), not code safety. The security scan covers known CVEs and suspicious install scripts.

Tools

9 tools · none privileged
Statically extracted from the published packagev0.1.3 · 20h ago

Read out of the source npm actually ships, at scan time. The package was never executed. Tools registered dynamically at runtime, or hidden inside bundled or minified code, can be missed — so this is a floor on the tool surface, not a complete census of it.

list_endpointsList the LLM endpoints currently monitored in this workspace.

List the LLM endpoints currently monitored in this workspace.

No input schema was published for this tool.

create_endpointRegister an LLM endpoint to monitor. The workspace must already have

Register an LLM endpoint to monitor. The workspace must already have

No input schema was published for this tool.

list_specsList behavioral specs in the workspace. A spec is a stored prompt +

List behavioral specs in the workspace. A spec is a stored prompt +

No input schema was published for this tool.

create_specCreate a behavioral spec. The first run after creation sets the

Create a behavioral spec. The first run after creation sets the

No input schema was published for this tool.

run_specRun a spec on demand and return the drift score immediately. Use this

Run a spec on demand and return the drift score immediately. Use this

No input schema was published for this tool.

reset_baselineClear a spec's baseline. The next run will record a new baseline

Clear a spec's baseline. The next run will record a new baseline

No input schema was published for this tool.

get_drift_eventsFetch recent drift events across the workspace, newest first. Each

Fetch recent drift events across the workspace, newest first. Each

No input schema was published for this tool.

get_spec_historyGet the run history for a single spec. Returns each run's status,

Get the run history for a single spec. Returns each run's status,

No input schema was published for this tool.

get_healthPer-spec health across the workspace: spec_id, spec_name, endpoint_name,

Per-spec health across the workspace: spec_id, spec_name, endpoint_name,

No input schema was published for this tool.

9 of 9 tools published a description.

Tool names and descriptions are written by the publisher and shown verbatim as inert text. They are the strings an MCP client passes to a model, so Forge scans them for prompt-injection patterns — any finding appears with the security scan above. “Privileged” is a keyword match on the tool name, not an audit of what the tool does: a benign-sounding name can still do anything.

About

MCP server for ModelWatch — continuous behavioral drift monitoring for LLM-powered applications.

Keywords
mcpmodel-context-protocolllm-monitoringmodel-driftai-observabilityllm-evaluationclaudeopenaidrift-detection
Alternatives
Comparing tool surfaces…

Tools exposed

9 tools · none privileged · observed by static extraction. A tool surface is what one scan read on one day, not a live capability list.

Linked names open Forge’s index of every entry observed exposing that tool. Browse all indexed tools.

Dependency tree

What one Forge scan resolved from npm metadata on 2026-10-11 — observed resolution, not a publisher declaration.

60 packages resolved · 1 direct · none carrying advisories Resolution stops at depth 4 and 60 packages.

The crawl stopped at the depth-4 limit. Anything below that level was never resolved.

The crawl stopped at the 60-package limit. The rest of the tree was never resolved.

36 more resolved packages are not drawn here (display cap: 24). Every dependency carrying an advisory is drawn regardless of the cap. Full inventory (CycloneDX SBOM)

Declared but not resolved

54 declared dependencies never landed in the tree. They are missing from Forge's resolution, not from the package.

+42 more not listed. The counts by reason above cover all of them.

Not followed: peerDependencies. This tree covers runtime dependencies only, so anything those pull in was never resolved.

Topics

Related in monitoring & logging