Now open source — Apache 2.0

Multi-model
code pipelines
with a referee

Every model has blindspots. CRTX routes your task through specialized stages — each handled by the best model — with an independent Arbiter that catches what they miss.

Start building →View documentation
Architect
Gemini 2.5 Pro
Arbiter
Grok 4
⚑ FLAG
⟨⟩
Implement
GPT-4o
● live
Refactor
Claude Opus 4.6
997
tests passing
5
pipeline modes
4+
LLM providers
$0.50–3
per pipeline run
~2–10min
average completion

See it run

Debate mode. Two models argue. A judge decides. The Arbiter reviews.

crtx — ~/project

How it works

01

You describe the task

Natural language. What you'd tell a senior developer. The CLI or dashboard accepts your intent and configures the pipeline.

crtx run "..." --arbiter bookend
02

Smart routing picks the best models

Fitness scores match each stage to the strongest available model. Four strategies let you optimize for quality, cost, speed, or a custom balance.

quality_first │ cost_optimized │ speed_first │ hybrid
03

Each stage builds on the last

Architect scaffolds, Implementer writes code, Refactorer improves, Verifier validates. Each model focuses on what it's best at. Cross-stage context flows automatically.

sequential │ parallel │ debate │ review │ improve
04

The Arbiter catches what they miss

An independent model — never the same one that did the work — reviews each stage. It can approve, flag warnings that get injected into the next stage, reject with retry, or halt the pipeline entirely.

APPROVE │ FLAG │ REJECT │ HALT
THE ARBITER

The model that watches the models

The Arbiter is an independent referee that reviews every pipeline stage. It uses a different model from the one that produced the work — enforcing cross-model accountability. When it flags an issue, structured feedback is injected into the next stage automatically.

This is the core insight: no model should grade its own work. The same way code review works on engineering teams — someone else always checks.

APPROVE
Output meets all criteria. Pipeline continues.
FLAG
Non-blocking warnings. Injected into the next stage as context.
REJECT
Stage retries with specific feedback. Max 2 retries before escalation.
HALT
Critical issue detected. Pipeline stops. Human review required.

Five modes for different problems

Not every task needs the same approach. Match the mode to the problem.

One stage at a time. Each builds on the last.

The default mode. Architect designs, Implementer codes, Refactorer improves, Verifier validates. Each stage inherits the previous output. The Arbiter reviews at configurable checkpoints — after the first stage, the last, both, or all.

crtx run --mode sequentialavg. cost: ~$0.50–3.00
BEST FOR Standard development tasks, feature implementation, bug fixes

Presets — one flag, zero config

Pick a preset and go. Each one configures mode, routing, arbiter, and model selection.

PresetModeCostTimeReviewUse case
balancedsequential~$0.754 minbookendDefault for most tasks
fastsequential~$0.302 minoffQuick iterations
cheapsequential~$0.403 minoffBudget-conscious
thoroughsequential~$1.505 minfullProduction code
exploreparallel~$2.004 mincross-reviewMultiple approaches
debatedebate~$3.008 minfullArchitecture decisions
crtx run "Your task" --preset balanced

Mode comparison

SEQUENTIALPARALLELDEBATE
Best forStandard dev tasksExploring approachesArchitecture decisions
Models used1 per stage3+ simultaneously3+ adversarially
Typical cost$0.50-1.00$1.50-2.50$2.50-3.50
Typical time3-5 min3-5 min6-9 min
Quality ceilingGoodVariableHighest
ArbiterOptionalRecommendedRequired

Three minutes to your first pipeline

Bring your own API keys. CRTX handles the rest.

quickstart.sh
# Install
pip install crtx

# Configure (bring your own API keys)
export ANTHROPIC_API_KEY=sk-ant-...
export OPENAI_API_KEY=sk-...

# Run your first pipeline
crtx run "Build a user auth service" \
  --arbiter bookend \
  --route quality_first

"We use CRTX to ship a production platform with 2,900+ tests (and CRTX's own 997). The Arbiter has caught schema mismatches, missing error handlers, and security gaps that single-model workflows missed entirely."

TriadAI·Makers of CRTX

Pricing that scales with you

Free CLI forever. Dashboard for teams. Cloud API for infrastructure.

OPEN SOURCE
Free
Apache 2.0 — forever
Bring your own API keys
Full CLI with Rich terminal output
5 pipeline modes including Review & Improve
Arbiter with 4 verdicts + retry logic
Smart routing engine (4 strategies)
Anthropic, OpenAI, Google, xAI
Custom domain rules via TOML
Unlimited local pipeline runs
Community support
Auto-fallback on provider outages
Apply mode — write code directly to your repo
Streaming real-time pipeline display
Context injection from your codebase
REPL mode for interactive sessions
pip install crtx
POPULAR
PRO
$29/mo per seat
14-day free trial
Your API keys · Our dashboard
Everything in Open Source
Web dashboard with 5 view modes
Session history & pipeline replay
Split comparison with model voting
Usage analytics & cost tracking
Team sharing & collaboration
Priority routing configuration
Email & Slack support
Start free trial →
CLOUD
$99/mo
Includes 15 pipeline runs
Managed keys · Hosted API · $7/additional run
Everything in Pro
Hosted API endpoint
Managed model API keys
No provider accounts needed
15 pipeline runs included
Usage dashboard & billing
Webhook notifications
99.9% uptime SLA
Priority support
Get API access →
Enterprise
Volume discounts, SSO/SAML, dedicated infrastructure, custom model integrations, on-prem deployment.
Contact sales →

Free and Pro tiers use your own API keys — you pay model providers directly.
Cloud tier includes all model costs. No separate API accounts needed.

Stop trusting one model
with your codebase

Any model. Any role. The protocol is the product.

pip install crtxView on GitHub →