Skip to main content
Glama

RealRun MCP server

Listed on mcpservers.org

Read-only lookups over RealRun's published, independent tests of AI models on everyday tasks (inbox triage, meeting scheduling, receipts, small-business finance, customer support replies). It lets an agent ask "which model should I use for task X?" and get scores and safety warnings back.

Connect (remote, no install)

Streamable HTTP endpoint: https://realrun.baseledgerip.com/mcp

{ "mcpServers": { "realrun": { "url": "https://realrun.baseledgerip.com/mcp" } } }

Claude Code: claude mcp add --transport http realrun https://realrun.baseledgerip.com/mcp

Cline: add this to cline_mcp_settings.json (keep "type": "streamableHttp"; without it Cline falls back to SSE and the connection fails). Full agent-readable steps: llms-install.md.

{ "mcpServers": { "realrun": { "type": "streamableHttp", "url": "https://realrun.baseledgerip.com/mcp", "disabled": false, "autoApprove": [] } } }

Related MCP server: Kryve Agent Evaluation MCP

Run locally (stdio)

npx -y realrun-mcp

(once published to npm; from a checkout: npm install && node src/stdio.mjs)

Tools

Tool

Input

Returns

list_tasks

none

Tested tasks, item counts, dates, results pages, playbook prices

get_scores

task

Every model's headline score (best first), cost tier, metric definition

recommend_model

task, priority (accuracy/cost/balanced), include_subscription

One pick + 2 alternatives; never picks a model flagged not recommended

get_warnings

task

Not-recommended and caution flags (scam failures, invented data, inconsistency)

Task ids: inbox-triage, meeting-scheduling, receipts, finance, support-replies.

All tools are read-only. The server stores nothing and has no auth. It is stateless, with JSON responses.

Data

src/data.json holds only the published headline numbers. Full prompts and item-level results are in the paid playbooks ($29 each, links in list_tasks). Scores reflect our tests on the dates shown, not a guarantee.

Billing (not active)

Every call passes through src/metering.mjs. The plan is a free allowance of about 20 lookups a day, then per-call billing via Stripe or x402. It is off: nobody is charged, and no payment objects exist.

Test

npm test runs stdio end to end. node test/smoke.mjs https://realrun.baseledgerip.com/mcp tests the remote endpoint.

Layout

  • src/server.mjs: tools

  • src/stdio.mjs: CLI

  • src/http.mjs: exports for the HTTP function

  • ../netlify/functions/mcp.mjs: Netlify Function at /mcp

  • server.json: official MCP registry metadata

License: MIT. © Base Ledger IP LLC.

Available Tools

4 tools
get_scoresGet model scores for a taskA
Read-only
Inspect

Headline scores for every model RealRun tested on a task, best first, with cost tier and the metric definition.

Inputs: task (required).

get_warningsGet safety warnings for a taskB
Read-only
Inspect

Models RealRun flagged as not recommended or needing caution on a task (scam failures, invented data, inconsistency).

Inputs: task (required).

list_tasksList tested tasksA
Read-only
Inspect

List the everyday tasks RealRun has tested AI models on, with item counts, test dates, results pages and playbook prices.

Inputs: no named parameters.

recommend_modelRecommend a model for a taskA
Read-only
Inspect

Recommend which AI model to use for a task given a priority: 'accuracy', 'cost' or 'balanced'. Never recommends models RealRun flagged as not recommended.

Inputs: task (required), priority, include_subscription.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool has a clearly distinct purpose: listing tasks, getting detailed scores, getting recommendations, and getting warnings. There is no overlap in functionality; an agent can easily select the right tool based on the need.

    Naming Consistency5/5

    All tool names follow a consistent verb_noun pattern: list_tasks, get_scores, recommend_model, get_warnings. The convention is predictable and uniform.

    Tool Count5/5

    With only 4 tools, the server is well-scoped for its purpose of providing AI model evaluation and recommendations. Each tool earns its place and the count is neither too thin nor too heavy.

    Completeness5/5

    The tool set covers the full lifecycle: discovering tasks (list_tasks), comparing models (get_scores), getting a tailored recommendation (recommend_model), and checking warnings (get_warnings). There are no obvious missing operations for this domain.

    Maintenance

    ActivitySlowing
    ResponsivenessLimited data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI assistants to retrieve sourced, dated data on consumer robots and physical AI: full-text search of robot sheets, side-by-side comparisons, and lookups of claims, prices, availability, capabilities, programmability and recent changes, with every value traceable to an evidence-graded source. Read-only and free with no key required, served over Streamable HTTP.
      1
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables AI agents to look up facts about notable machine-learning models — organization, release date, parameter count, training-compute (FLOP) estimate, hardware and accessibility — by name, and to search or aggregate them by organization, domain, year, or compute threshold. Supports filtering the dataset newest-first and producing per-year training-compute trends, keylessly and without authentication.
      126 npm
      MIT