The AGI Scorecard — MCP server & dataset

Every prediction in Leopold Aschenbrenner's Situational Awareness (June 2024), graded against reality with pre-registered flip conditions, plus one auditable 0–100 score for the whole 2027 bet. Listed in the official MCP Registry as com.agiscorecard/agi-scorecard.

Thesis Tracker: 62.5/100 · verdicts as of 2026-09-06

This repository is a mirror. The single source of truth is agiscorecard.com. sync.mjs rebuilds the dataset tables from the live source every day; the integration notes and related resources below are curated.

MCP server (connect an agent directly)

claude mcp add --transport http agiscorecard https://agiscorecard.com/mcp

{ "mcpServers": { "agiscorecard": { "type": "http", "url": "https://agiscorecard.com/mcp" } } }

No auth, no install, streamable HTTP. Setup notes: https://agiscorecard.com/for-agents

4 tools

Related MCP server: analysis-gym

Why this dataset is different

• Pre-registered flip conditions. Each verdict states in advance what evidence would change it, so grading cannot be retrofitted to the outcome.

• The misses stay published. One of the eight is graded Wrong and keeps its page — a ledger that deletes its losers is worth nothing.

• One number, recomputed from the verdicts. The Tracker is not an opinion: supportive = 1, open/pending = 0.5, refuted = 0, averaged. Full history below.

• Dated. Every figure carries the date it was last reviewed.

Current verdicts

| 🟢 On track | Models outpace college graduates across knowledge work | 2025/26 | detail | | 🟢 On track | Compute + algorithmic scaling continues at trend | ~0.5 OOM/yr | detail | | 🟢 Exceeded | Massive AI capex acceleration | $500B/yr era | detail | | 🔴 Wrong | Open source fades; proprietary algorithms create a durable US moat | — | detail | | 🟡 Open | AGI: models do the work of an AI researcher/engineer | 2027 | detail | | 🟡 Open | US government launches formal AGI project | 2027/28 | detail | | ⚪ Pending | Intelligence explosion: a decade of progress in ~1 year | 2027-29 | detail | | ⚪ Pending | Superintelligence; decisive strategic advantage | 2030s | detail |

Thesis Tracker history

| 2026-06-30 | 62.5 | | 2026-07-12 | 62.5 | | 2026-08-08 | 62.5 | | 2026-09-06 | 62.5 |

Machine-readable

Also served live: /llms.txt · /data.json · /index-history.json

Related local review tools

The same team maintains Web3 Workbench, a free collection of local worksheets for AI and Web3 operational review:

• Agent Evidence: worked scenarios — inspect deduplicated task evidence, version boundaries and conflicting reviews.

• Route Lab: worked scenarios — compare a fallback's expected cost with its worst-case budget and latency.

• Protocol Ledger: methods and API — read dated protocol profiles with official sources and explicit scope limits.

These tools use a separate service from the dataset MCP above. Browser worksheet inputs are processed locally; example results are fictional. No wallet connection, model call, production routing or security certification is provided. See each tool's method, limits and separate software license.

Web3 Workbench MCP

Connect to https://web3.agiscorecard.com/mcp using Streamable HTTP, without an API key. Setup and available tools.

The endpoint exposes ten deterministic calculators plus search and fetch for public methods. Results include a canonical method citation, release version, sources and limitations. Each tool subdomain also provides a scoped /mcp endpoint. These are independent of the four dataset tools and registry identity above.

Remote MCP calls send the selected arguments to the server; application code does not persist inputs or results. Client and hosting-provider policies are separate. Public citations describe the method, not an independently verified private result. Connecting a client does not imply automatic inclusion in AI search or a registry.

Licence

Dataset: CC BY 4.0 — free to use with attribution to agiscorecard.com. See LICENSE.

Not investment advice; research and education only.