Brier

Forecasting as a harness for decision-making.

Brier is also the forecasting agent being developed inside Thesis: an open-source, agent-only lab for automatically resolvable public-data forecasts. For the current Thesis/Brier product vision, read docs/thesis-vision.md first. Future coding agents should start with AGENTS.md.

Instead of asking "Is X good?" or "Should I do Y?", brier helps you:

1. Define what success looks like (KPIs)
2. Expand your options (including ones you didn't consider)
3. Make explicit forecasts (with confidence intervals and resolution rules)
4. Track outcomes to improve calibration over time

Installation

python -m pip install 'brier[mcp] @ git+https://github.com/ThesisInstitute/thesis'

This installs from this repository. The brier project on PyPI belongs to someone else and is unrelated, so pip install brier fetches a different package. pip checks out the whole repository to build the package: about 2 GB, most of it forecast records.

Quick Start

Codex

brier setup codex
brier doctor codex

Then restart Codex and use $brier when a decision prompt appears.

Claude Code

brier setup claude
brier doctor claude

Then restart Claude Code.

Local CLI

brier new "Should we rewrite the auth layer?" --context "3 incidents this quarter"
brier list
brier calibration

The CLI is local-only and does not call an LLM or require an API key.

Python package

from brier import Decision, KPI, Option, Forecast, DecisionStore
from datetime import datetime, timedelta

# Create a decision
decision = Decision(
    question="Should I take the new job offer?",
    kpis=[
        KPI(name="income", description="Total comp after 2 years", unit="$"),
        KPI(
            name="satisfaction",
            description="Job satisfaction 1-10",
            outcome_type="score",
            resolution_date=datetime.now() + timedelta(days=365),
            resolution_rule="Ask for a 1-10 retrospective self-rating 12 months after starting.",
            data_source="Follow-up self-review",
        ),
    ],
    options=[
        Option(
            name="Take new job",
            description="Accept the offer at Company X",
            forecasts={
                "income": Forecast(
                    point_estimate=300000,
                    confidence_interval=(250000, 400000),
                    reasoning="Base + equity, assuming normal vesting",
                ),
                "satisfaction": Forecast(
                    point_estimate=7.5,
                    confidence_interval=(6, 9),
                    reasoning="Interesting work, but unknown team",
                ),
            }
        ),
        Option(
            name="Stay at current job",
            description="Decline and stay",
            forecasts={
                "income": Forecast(
                    point_estimate=250000,
                    confidence_interval=(230000, 280000),
                    reasoning="Known trajectory, likely promotion",
                ),
                "satisfaction": Forecast(
                    point_estimate=6.5,
                    confidence_interval=(6, 7),
                    reasoning="Comfortable but plateauing",
                ),
            }
        ),
    ],
    review_date=datetime.now() + timedelta(days=180),
)

# Save it
store = DecisionStore()
store.save(decision)

Command Line

brier new "Should we launch now?"
brier show abc123
brier pending
brier calibration

Forecast Question Drafts

brier can turn a stored decision forecast or standalone policy question into Manifold-ready forecast question drafts. This is draft-only: it does not publish questions, place a bet, or require a Manifold API key.

brier forecast-draft "Will Waymo be legally permitted to offer fully driverless paid robotaxi rides in Washington, DC by 2026-12-31?" \
  --initial-prob 52 \
  --resolution-date 2026-12-31 \
  --resolution-rule "Resolve YES if official DC law, regulation, or permit approval allows Waymo to offer fully driverless paid public rides in DC by 2026-12-31." \
  --source "Waymo DC announcement|https://waymo.com/blog/2025/03/next-stop-for-waymo-one-washingtondc/" \
  --source "DC AV testing law|https://code.dccouncil.gov/us/dc/council/laws/23-156" \
  --tag waymo \
  --tag dc \
  --output waymo-dc-forecast-pack.json

For a stored decision with options and forecasts:

brier forecast-draft abc123 --output forecast-pack.json

An example Waymo/DC draft pack lives at examples/waymo_dc_market_pack.json. It uses an existing Manifold public-service question as the gate, then drafts conditional aggregate 2027 safety forecasts for DC traffic fatalities and serious injuries. These resolve N/A when the linked gate question resolves the opposite way.

AI Agent Workflows

brier is not tied to Claude. The Claude Code plugin is the most integrated path today, but the framework also works with Codex and other coding agents that can follow structured instructions or run shell commands.

For agent-agnostic setup and prompt guidance, see docs/agent-workflows.md.

Codex and other coding agents

The default builder path is package-first:

python -m pip install 'brier[mcp] @ git+https://github.com/ThesisInstitute/thesis'
brier setup codex
brier doctor codex

For source installs during development:

python -m pip install -e /path/to/brier

MCP server

If you want a native tool interface instead of prompt copy-paste, install the package and run the MCP server locally:

python -m pip install 'brier[mcp] @ git+https://github.com/ThesisInstitute/thesis'
brier-mcp

It exposes tools for creating, listing, retrieving, saving, and scoring decisions, plus resources/prompts for the brier workflow.

To register it in Codex as a local MCP server:

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗