hedwigai

에이전트를 위한 핵심 데이터: 구매자 목록, 주간 보고서, KPI 트래커와 같은 워크북으로, 스스로 조사하고 모든 출처를 인용하며 최신 상태를 유지합니다.

호스팅형 MCP 서버

npx add-mcp 'https://mcp.hedwigai.com'

Claude Code, Codex, Cursor 등에 설치됩니다

문서

Work you can check.

A system of record for the spreadsheets and documents your AI edits, checked before every save.

First workbook free. Reading and writing workbooks costs nothing.

Works in

Files stay in

Listed in the Claude Connectors Directory · MCP Registry · mcp.so, verified

Enterprise? A private model trained on your own data →

Which are you?

For Claude Code, Codex, Cursor and any MCP agent

Agent memory that outlives the session.

  • CLAUDE.md and memory files stay on one machine. Workbooks hold decision logs, task status, handoffs and results that any session, agent or teammate on your account can read.
  • Every write is a new version; a write made on an old version is refused, never merged.
  • Each write is checked against the workbook's schema, the way a compiler checks code, with an error for each bad cell.
  • An MCP server, connected in one line: claude mcp add --transport http hedwigai https://mcp.hedwigai.com

For people and teams

Let Claude or ChatGPT edit your Excel and Word files in place.

  • Your files stay in Google Drive or Dropbox, and each edit is written back there. No downloading and uploading again.
  • Every edit is checked against the file's schema before it is saved; an edit that breaks it is refused, with the cell that broke it.
  • Every version is kept; an edit made on an old version is refused, never merged.
  • Set up in the AI app you already use. The first workbook is free.

Try it now, nothing to install.

01 · MCP for knowledge work

A compiler for knowledge work.

Your spreadsheets and documents stay in your Google Drive or Dropbox; the file is the record. Every AI assistant, and every run it starts, reads and edits that file through one layer, the workbook. Its schema declares each table's columns, types, required cells, keys and references, so the tooling that keeps code correct now keeps your operating records correct.

> find 50 leads for our industrial pumps and draft an outreach email to each

Accurate

Every edit is checked against the schema before it is saved. A change an assistant proposes is checked on a copy first and redone with the errors, so you are only asked to approve one that passes.

Resilient

Every version is kept and can be restored. An edit based on an old version, or over a file someone changed in Excel meanwhile, is rejected, never merged over.

Fast and cheap

An assistant reads and edits one table or section by its address, not the whole file, so a change costs only what it touches.

Works inside Claude, ChatGPT, Codex and Cursor.

  • Claude
  • Codex
  • Cursor
  • Any MCP app

A short answer from your workbooks and programs, with the source it rests on. Nothing to wait for.

Is Moderna a biotech with a public float above $5B?

Your agent

ask · Is Moderna NASDAQ:MRNA a biotech company with public float above $5B?

{
  "answer": "yes",
  "claims": [{ "figure": "public float", "value": "$9.6B",
               "as_of": "2026-06-28" }],
  "sources": ["MRNA 10-K, cover page",
              "biotech-diligence-7fq2 · §Moderna"]
}

Yes. Moderna's public float was $9.6B as of June 28, per the cover of its 10-K.

02 · Interpretability experiments

Why a model behaves as it does, measured inside it.

Interpretability explains a model's behaviour from its internals: which neurons and features respond to an input, and how the output changes as the input, the prompt or the environment changes. We test published claims on open-weight models with causal methods (ablation, activation steering, probes and weight edits), each against a control, and publish every number, including the claims that failed. You can rerun each one.

59 of 100

Does one neuron catch reward hacking?

The neuron posted as a reward-hacking detector in Qwen3.5-4B tells a hack from an honest answer in 59 pairs of 100. Another neuron, found with a held-out search, does it in 99.

Run it yourself →

9.2% vs 8.6%

Does emergent misalignment run through one neuron?

Switching off the neuron that makes the model hostile leaves 9.2% of answers misaligned. A random control neuron gives 8.6%. The claim does not hold.

Run it yourself →

91 of 100

Can a monitor reading the thinking catch a pushed neuron?

A Qwen3.5-9B monitor reading Nanbeige4.3-33B's chain of thought catches a full personality push in 91 pairs of 100, and a half push in 65.

Run it yourself →

22 vs 7

Does a weight edit stop a model looping?

On prompts with no math, 22 of 32 answers from the released Qwen3.5-0.8B ran to the 8,192-token limit. The edited copy did it 7 times.

Run it yourself →

Few example workbooks

VK2735 competitive landscape: the 8 closest obesity programs

Tirzepatide, CagriSema, retatrutide, orforglipron and four more against Viking's VK2735: phase, latest weight-loss readout, next catalyst and deals, every row cited.

29 sources · 25 minutes

Read the workbook →

US importers of bulk black tea from India, from customs records

Four US buyers with the bills of lading that show it, who they buy from, what gates a new supplier, and the leads not worth chasing yet.

10 sources · 8 minutes

Read the workbook →

Early-stage AI startups that filed a Form D in the last 30 days

Seven US AI companies raising in September 2026: the instrument each filing names, amount sold, investors and the people on it, each cited to the Form D on EDGAR.

17 sources · 13 minutes

Read the workbook →

Every sample

Tested on real work. Every score published.

Can a 32M-parameter model parse references better than GROBID?

On 700 real Crossref references, citeparse-150m got 584 fully right, citeparse-32m 562 and GROBID 0.8.1 498. GROBID still wins on its own held-out citations.

citeparse-150mciteparse-32mGROBID 0.8.1
References fully right, of 700584562498
Fields right, wild set0.9520.9390.911

Read the benchmark →

Which assumption moves a walking robot's score most in simulation?

Stiffness inside the range a biped's walking policy was trained on moves its turn score from 52% to 99%; new soles move it 4 points.

Value
Turn rate reached, stiffness ×0.8 to ×1.252% to 99%
Turn rate reached, latency 1 vs 2 steps68% vs 49%

Read the numbers →

Does Physical Agent aim its tool calls at the object?

Physical Agent writes a robot tool call on 400 of 400 frames, but moving the object 16 cm leaves its targets where they were.

Physical AgentQwen3.5-4B
LIBERO frames answered with a tool call400 / 40018 / 400
Target movement per cm of object movement (median)0.02 cm—

Read the numbers →

Frontier models by scale, 2021–2026: every compute band climbs at the same pace

Each band of training compute gains 12 to 20 capability points a year; a lab's smallest model trails the best by about seven months.

Value
ECI points gained per year, by compute band12 to 20
Delay for a band ten times smaller to match3 to 13 months

Read the numbers →

Particle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash in Claude Code with hedwigai workbooks

Particle-1, DeepSeek V4.1 Flash and GLM-5.3-Flash, 180 runs graded blind, rows right: 91.2%, 89.4% and 90.8%, intervals overlapping.

Particle-1DeepSeek V4.1 FlashGLM-5.3-Flash
Rows right, blind91.2%89.4%90.8%
Median seconds per run375315144

Read the numbers →

Who imports it? Four workbooks that keep the answer, from US customs records

Marin Mountain Bikes filed more bicycle shipments than Giant or Trek; 572 companies imported black tea.

US customs records
Most bicycle shipmentsMarin Mountain Bikes, 5,711
US black tea importers, 2026572

See who imports it →

All 54 experiments, every score published

Questions, answered.

What does hedwigai do?

It is a system of record for the spreadsheets and documents an AI edits. Claude, ChatGPT, Codex, Cursor or any MCP agent keeps them as workbooks, optionally linked to a file in your Google Drive or Dropbox. Every write is a new version, a write made on an old version is refused rather than merged, and every edit is checked against the workbook's schema before it is saved.

Why keep an agent's state in hedwigai rather than a local file or its memory?

A local file stays on one machine and is overwritten by whoever writes last; an agent's memory is private to it. A workbook is found by the next session and by other people and agents, keeps every version, refuses a write based on an old version, and holds the shape its schema declares, so a tracker or a decision log does not drift.

How do I connect my agent?

In Claude Code: claude mcp add --transport http hedwigai https://mcp.hedwigai.com, then /mcp to sign in, or install the hedwigai plugin with /plugin marketplace add un-seen/hedwigai-skills. In Claude, add it from the Connectors Directory. Codex, Cursor, Hermes and any other MCP app add https://mcp.hedwigai.com as a server.

Where do my files live if I connect the MCP?

In your Google Drive or Dropbox, at the same path. The file is the record: a run reads the latest version first and writes each new version back to it. If someone changed the file in the meantime, the edit is refused rather than written over theirs, and every version can be restored.

Which AI apps does it work with?

Claude, where it is listed in the Connectors Directory; ChatGPT, added as a custom MCP server; and Claude Code, Codex, Cursor, Hermes or any other app that adds an MCP server. It is also listed in the MCP Registry, on Glama and on mcp.so.

What does it cost?

The first workbook is free. Reading, searching and writing workbooks is free; research runs, started only when you ask for research, are billed to the account by what they use.