NoiseFloor

LIQUILENS PRIVATE LIMITED / beepboop2025 · Data & Analytics Infrastructure

MCP ServerCommunity

Statistical research diagnostics for checking whether financial changes are signal or noise. Ten tools assess caller-supplied market observations, repeated headlines, change and sampling effects, empirical forecast intervals, forecast performance and correlation structure. MIT-licensed Python toolkit with local and hosted MCP access; does not fetch market feeds, retain portfolios or execute trades.

statistical diagnosticschange detectionforecast evaluationdata qualitycorrelation analysisheadline triage

Status

Community

Auth

None

Hosting

Both

Community rating

Not rated yet

Role

Domain

Specialized public intelligence. Turns a public dataset into a queryable, agent-ready domain expert. More on Domain servers → All seven roles →

Open documentation →

Caveats

Independently checked on October 13, 2026: the hosted endpoint answered an MCP handshake as version 0.4.0 and exposed ten tools without an API key. Statistical performance and the vendor's privacy claims were not independently verified; the local package was not executed.
Works on caller-supplied observations and headlines, not a market feed or portfolio system. Change statistics are descriptive, empirical forecast coverage is not a guarantee, and findings do not establish predictive alpha, investment suitability or causation.
which_metrics_matter requires the caller to supply valid e-values and attest valid_evalues=true; arbitrary scores are not valid inputs. spectral_assessment is descriptive correlation research, while dyson_reference generates a synthetic reference, not observed market evidence.
Local stdio use needs no account or API key. Hosted use sends supplied inputs to the operator; prefer local execution for sensitive client data. The repository says it does not retain portfolios and HTTP logs exclude observations and titles, but this is not an independent privacy audit. Self-host HTTP behind your own TLS, authentication and quota controls.
MIT-licensed source and pinned local MCP configuration are documented at https://github.com/beepboop2025/noisefloor. Package releases and hosted versions can differ, so check the live version and tool list. Headline grouping prioritizes review and does not certify truth.
Start with read-only scopes and keep client-side approval prompts on until the connection is proven.

Rate this server

How to connect NoiseFloor

Step 1

Remote server: paste a URL

NoiseFloor publishes a remote (Streamable HTTP) endpoint, so nothing is installed locally. Add it as a custom connector in Claude, ChatGPT or your IDE and paste the endpoint below.

Endpoint

https://api.seiche.info/noisefloor/mcp

Client config snippet

{
  "mcpServers": {
    "333a058e-c776-4c86-b0f5-980ed0b809b0": {
      "url": "https://api.seiche.info/noisefloor/mcp"
    }
  }
}

Step 2

Authenticate with no credentials

No credentials are required for public data. Rate limits still apply, so avoid bulk pulls from a chat client.

Step 3

Verify, then widen access

Restart your client, confirm the server's tools appear, and run one read-only question first. Keep client-side approval prompts on for any tool that can write, trade or send. Confirm with your firm's compliance policy before connecting systems holding client data.

Prompts that use this server

Starter prompts from the Practice Stacks that rely on Data & Analytics Infrastructure.

Custom reporting query brief

A firm-level answer pulled from your warehouse.

Portfolio Reporting & PerformanceData & Analytics InfrastructureCustody & Clearing

Read-only: answer <question, e.g. AUM by custodian and household size> from my analytics warehouse, then spot-check the answer against the live custodial data for two or three sample accounts. Show the SELECT query you ran, the row counts, the result table, and a reconciliation line comparing warehouse totals to custodian totals with any variance called out. Use SELECT only. No inserts, updates, deletes or schema changes, and do not place or modify anything at the custodian. If a table or account you need is missing, say so instead of approximating.

Guardrail · Read-only credentials only. Never connect a warehouse with write access to a chat tool.

Warehouse schema map

A plain-English map of what data you actually have.

Data Infrastructure (Builder / Enterprise)Data & Analytics InfrastructureWealth Platforms

Read-only: map the warehouse schemas and tables relevant to <domain, e.g. client holdings>, table, purpose, grain, row count and how tables join, then trace each one back to the upstream wealth platform or custodial API field that feeds it. Output a lineage table of Warehouse table.column / Upstream source field / Grain / Last loaded, and flag tables with no identifiable upstream source or with fields the source no longer provides. Use metadata and SELECT queries only. No DDL, no writes, no schema changes.

Guardrail · Read-only credentials. Never connect a warehouse with DDL rights to a chat tool.

Data quality profile

Nulls, duplicates and outliers in the tables you rely on.

Data Infrastructure (Builder / Enterprise)Data & Analytics InfrastructureMarket & Financial Data

Read-only: profile <table or dataset>, null rates by column, duplicate keys, value ranges and obvious outliers, then validate a sample of its prices, identifiers or corporate-action fields against a market-data source for the same dates. Show the SELECT queries you ran, the results, and a comparison table of warehouse value vs. market-data value vs. variance. Do not write, delete or correct any data. Recommend fixes as a list for my engineer to review.

Guardrail · Read-only. Corrections belong in your pipeline, with review and version control.

Read the full connection guide →