Hokusai

Rework Risk · Early access

Find out what happened to the code your agents merged.

Rework Risk scans your merged PRs and shows which held up, which needed fixing within 30 days, and what predicted it.

Tests tell you code passes. Your Rework Report shows what happened after it passed.

Or start routing coding tasks today. Model Match is live.

What's in your Rework Report

Measured from your own history.

Every merged PR is checked against what happened next: was it reverted, fixed, or tied to an incident? Your git history already records the answer.

7, 30 and 90 days

Rework measured at three horizons after merge.

3 outcome signals

Reverts, fix PRs, and incidents linked to the change.

What predicted it

Which kinds of change, by size, tests touched, and agent or human author, needed fixing most often in your repo.

Agent decisions, informed by real outcomes.

Rework Risk

Early access

Review the agent code that needs it.

A 30-day rework risk score on every pull request, learned from what actually happened across many codebases.

See how it works

Model Match

Live

Spend less on models without losing quality.

Picks a model from your pool for each coding task, based on how similar tasks turned out.

Start routing

Your decision here

Open to partners

Bring outcome data for a new model.

Tool selection, escalation, and retrieval are next. Have outcomes for a decision your agents make? Build it with us.

Talk to us

Hokusai is a protocol for setting up decision models for AI agents and the incentives around them. Each model stands on its own and learns from outcomes across many organizations. Hokusai measures every improvement and gives the people who made it a stake in the model.

How it works

Each model sits beside your agents as an advisor. It learns from what happened, not from whether work looked good.

  1. 1

    Your system asks a model

    Your agent or pipeline asks about a decision: how risky is this change? Which model should take this task?

  2. 2

    The model predicts

    Rework Risk returns a 23% rework risk. Model Match returns a recommended model. Each is learned from outcomes across many organizations.

  3. 3

    You decide

    Your system decides what to do with it. No model blocks a merge or proxies a model call.

  4. 4

    Outcomes flow back

    What actually happened (a revert, a fix PR, a completed task) improves the model, and Hokusai rewards whoever supplied the outcome.

Not a gateway

Our models never sit in your request path.

Not an eval tool

They predict what happens next.

Not a harness

You keep your agents and your workflow.

For agent builders

Keep your harness. Stop hand-building every decision inside it.

Whether you run Claude Code and Codex, a multi-agent harness, or custom agents built in-house, your loop is full of decisions: which model, whether a change is safe to rely on, where a human should look. You can train those on your own history, but the outcomes that matter are rare and slow to show up. Models built on Hokusai learn from many teams at once.

Neutral

No preferred harness, model provider, or tool.

Advisory

A score or recommendation your code acts on, outside the request path.

Owned

Keep the ownership your outcomes earn, or pass it through to your users.

Integration

Bring Hokusai to your existing workflow.

Use Hokusai from coding assistants, agent harnesses, backend services, or direct API calls. Pick the path that matches your workflow, then route tasks and report outcomes.

1

Choose an integration point

Start from the workflow you already use: a coding assistant, an agent harness, application code, or a direct API call.

2

Route against your model pool

Pass only the candidate models your environment can actually run. Model Match recommends one; your system still executes the task.

3

Report outcomes

Send coarse success, cost, latency, and token signals back so Model Match improves.

Claude CodeCodexWavemillTypeScriptPythonREST API
See all integration paths

Coding / Multi-Model Routing

Model Match

Model Match's first training corpus comes from real autonomous coding evals, not synthetic routing examples. The data spans migrations, features, bug fixes, infra, tests, refactors, and docs across TypeScript, JavaScript, Python, Bash, and mixed-language repos. This provides a starting point for learning routing behavior while leaving clear room to expand coverage across more repositories, languages, and task families.

How Model Match works

When we're compared to…

Eval platforms

Braintrust, LangSmith

They help you evaluate what your agents did. Rework Risk learns what happened afterward, across organizations, and predicts which work will hold up. They work well together.

AI code review

Review bots, static analysis

They review the PR as it looks today. Rework Risk predicts what happens to it after merge, scored against what actually happened.

Tests

Unit, integration, CI

Tests tell you code passes. Rework Risk predicts what happens to code that passed.

Routers and gateways

OpenRouter, Not Diamond

Model Match is a model router that never sits in your request path. It recommends a model from your pool based on how similar tasks turned out.

Building it in-house

Your own history

You can. But rework is rare and slow to observe, so no single team sees much of it. Rework Risk learns from far more, and you keep your harness.

Shared intelligence. Shared ownership.

Hokusai is the protocol underneath every model. Each model is owned by the people who improve it. When your outcome data produces a measurable improvement, the protocol mints you tokens in that model. For Model Match, that token is HROUT.

Integrators pay per-decision fees in USDC, and tokens redeem against those fees. Keep your stake, pass it through to your users, or redeem it for USDC anytime.

How ownership works