verd

Five minds enter. They argue, challenge, cross-examine. Only the truth walks out.

verd spawns multiple AI models from different families — each with a specialized role, has them debate your question across rounds, then a stronger judge delivers the final verdict with strengths, issues, and actionable fixes.

Use it everywhere: CLI for code reviews, MCP inside Claude Code and Cursor, and Slack as @verd in any conversation.

Getting Started

Requires Python 3.11+.

pip install verd
verd setup

The setup wizard walks you through provider selection (OpenRouter, LiteLLM, or other) and outputs the exact config you need — for both CLI (.env) and MCP (JSON to paste into your editor config).

verd runs multiple models in parallel (Claude, Gemini, GPT, DeepSeek) so it needs a multi-provider router. OpenRouter is the easiest — one key, all models. LiteLLM proxy works too.

Usage

CLI

verd "can this auth middleware be bypassed?" -f auth.py middleware.py
verdh "should we merge this?" -gb main         # deep mode — 5 models, 3 rounds

MCP (Claude Code / Cursor) — use verd, verdl, verdh as tools directly in chat:

verdh based on the context above and this file, do you think we can proceed?
is this approach correct given what we discussed, use verdl

Slack — mention @verd in any channel or thread:

@verd what do you think?          — reads thread context, debates, replies
@verd deep is this secure?        — uses verdh (5 models, 3 rounds)
/verd Kafka, SQS, or RabbitMQ for our event pipeline?  — slash command with live progress

verd is for critical decisions and deep analysis — not simple lookups. If a single model can answer it, verd is overkill.

Output

FAIL  77%  In-memory rate limiter is unsafe for production

claude:FAIL  gpt:FAIL  gemini:FAIL  gpt:FAIL  (FULL)

+ Conceptually correct sliding-window logic
- Global dict is unsynchronized — race conditions in multi-thread servers
- Per-user lists grow without bounds — memory leak / DoS vector
! gpt-5-mini caught the risk of system clock jumps with time.time()
→ Move state to Redis with atomic operations

completed in 69.3s • 22,449 tokens • ~$0.07

Vote breakdown, unique catches (!), dissent, strengths, issues, and actionable fixes — all in one view.

Modes

Benchmark

Tested on the Martian Code Review Benchmark — 50 real PRs from Cal.com, Discourse, Grafana, Keycloak, and Sentry with expert-labeled golden comments. No code-review-specific tuning.

+37% F1 over Claude solo. 57% more precise. 42% fewer false positives.

How it works

1. Your question + content gets sent to multiple AI models in parallel
2. Each model has a specialized role (analyst, devils_advocate, logic_checker, fact_checker, pragmatist)
3. Models see each other's responses and cross-examine for 1-3 rounds
4. Anti-groupthink prompts ensure models hold their ground when they have evidence — consensus without new evidence is rejected
5. A stronger judge model synthesizes the debate, weighting each reviewer by their role
6. Confidence is calculated from vote distribution — a fact_checker's dissent lowers confidence more than a devils_advocate's expected pushback
7. You get: verdict, vote breakdown, strengths, issues, unique catches, dissent, and actionable fixes

The key insight: different model families have different blind spots and training biases. Claude spots nuance GPT misses. Gemini catches logic errors DeepSeek overlooks. More importantly — if the same model writes the review and judges its quality, it's likely to agree with itself. Cross-model diversity means the judge is a genuine quality gate, not a model grading its own homework. The debate surfaces what each model uniquely caught and tells you exactly which model caught what.

Roles

The judge weighs each reviewer's input by role — a fact_checker citing sources carries more weight than a devils_advocate pushing back.

Config

Override models via env vars or CLI flags. Per-tier env vars let you set different models for each mode:

VERDL_JUDGE=o4-mini            VERDL_DEBATERS=gpt-4.1-mini,gemini-3.1-flash-lite-preview
VERD_JUDGE=o3                  VERD_DEBATERS=claude-sonnet-4-6,gpt-4.1,gemini-3.1-pro-preview,gpt-4.1-mini
VERDH_JUDGE=o3                 VERDH_DEBATERS=claude-opus-4-6,deepseek-r1,gemini-3.1-pro-preview,sonar-pro,gpt-4.1

Or use VERD_JUDGE / VERD_DEBATERS as a global override for all tiers. verd setup generates the right config for your provider.

Flags

-c TEXT               inline content string
-f FILE [FILE ...]    one or more files to evaluate
-d [DIR]              read all files in a directory (default: current dir)
-g                    use unstaged git diff as content
-gs                   use staged git diff as content
-gb REF               use git diff REF...HEAD as content (e.g. main)
-a / --all            scan all files, skip smart selection (use with -d)
--ext EXT [EXT ...]   filter by extension (use with -d)
--exclude PATTERN     glob patterns to exclude (use with -d)
-q / --quiet          hide debate transcript, show only verdict
--json                output raw JSON
--judge MODEL         override judge model
--debaters MODEL ...  override debater models
--budget USD          max cost in USD — abort if estimate exceeds budget
--timeout SECONDS     override timeout per model call
--version             show version and exit

MCP — Claude Code / Cursor

verd setup    # select "MCP" and your provider

This prints the exact JSON to paste into ~/.claude/settings.json (Claude Code) or ~/.cursor/mcp.json (Cursor), with the correct absolute path to verd-mcp and model overrides for your provider. Then use verd, verdl, or verdh as tools directly in chat.

Slack

Install with Slack dependencies:

pip install "verd[slack]"

Create a Slack app with Socket Mode enabled, add bot scopes (app_mentions:read, channels:history, groups:history, chat:write, reactions:write, im:history, im:write, users:read), then:

export SLACK_BOT_TOKEN=xoxb-...
export SLACK_APP_TOKEN=xapp-...
export SLACK_SIGNING_SECRET=...
verd-slack

Optional: restrict access via environment variables:

export VERD_ALLOWED_CHANNELS=C123,C456    # empty = all channels
export VERD_ALLOWED_USERS=U123,U456       # empty = all users