crosscheck-ai

AI Dev Team -- multi-agent coding pipeline with 9 specialized agents, 8-phase workflow, and real-time security monitoring.

Built on OpenRouter. 36 verified models. Free by default. 283 tests passing.

What is crosscheck-ai?

crosscheck-ai is a team of 9 AI agents that work together like a real dev team -- and it's free to start:

Qwen 3 Coder (Coder) -- writes production code (480B MoE, free)
Qwen 3.6 Plus (Architect) -- designs system structure (1M context, free)
Nemotron 3 Super (Planner) -- CEO, decomposes tasks (120B, free)
Nemotron 3 Super (Debugger) -- hunts bugs and edge cases (free)
Qwen 3 Next (Security) -- scans for vulnerabilities (free)
Trinity Large (Analyst) -- flow and performance analysis (free)
Qwen 3.6 Plus (Coordinator) -- safety gate, stress-tests consensus (free)
Nemotron 3 Super (Observer 1) -- adversarial reviewer (free)
Hermes 3 Llama 405B (Observer 2) -- diverse adversarial reviewer (free)

All default models are $0 cost on OpenRouter. You can swap any role to premium models (Claude, GPT-5, Grok, etc.) with --models.

Features

Feature	Description
9-Agent Dev Team	Coordinator + Planner + Architect + Coder + Debugger + Security + Analyst + 2 Observers
8-Phase Workflow	PLANNING -> DISCUSSION -> CODING -> REVIEW -> DISSENT -> APPROVAL -> TESTING -> DONE
Free by Default	All 9 agents use free OpenRouter models -- $0 cost to start
InsAIts Security	Real-time AI-to-AI monitoring: catches prompt injection, credential leaks, jargon drift
ProjectWorkspace	Safe file operations with change proposals, rollback, and path traversal prevention
Group Chat UI	Web-based dialog panel -- talk to all agents at once via WebSocket
@Mentions	Route tasks to specific agents: `@coder fix line 45`, `@security scan for vulns`
Cowork Mode	Watch your project while you code in Cursor/VS Code/Claude Code
Code Review	Multi-agent review with dual-supervisor voting and conservative reconciliation
Observer Mode	Watch external coding sessions, flag bugs/security/regressions in real-time
20+ Languages	Auto-detect user language, all agents respond in the same language
36 Models	Full OpenRouter registry -- any model can fill any role
Local Models	Ollama/LM Studio support -- zero API cost with local models
Dashboard	Web UI for session history, cost trends, model performance
Tool Sandbox	Auto-run pytest, ruff, mypy, bandit after code changes
VS Code Extension	Inline decorations, CodeLens, live observer

Install

Basic install (CLI + team + review)

pip install crosscheck-ai

With Group Chat UI (recommended)

pip install "crosscheck-ai[chat]"

Full install (all features including InsAIts monitoring)

pip install "crosscheck-ai[all]"

For development

git clone https://github.com/Nomadu27/crosscheck-ai.git
cd crosscheck-ai
pip install -e ".[dev]"

Quick Start

1. Set your API key

crosscheck-ai uses OpenRouter, which gives you access to all 36 models with a single key. The default team uses only free models -- you won't be charged.

export CROSSCHECK_API_KEY=sk-or-v1-xxxxxxxxxxxxxxxx

On Windows:

set CROSSCHECK_API_KEY=sk-or-v1-xxxxxxxxxxxxxxxx

2. Launch the AI Dev Team (terminal)

crosscheck team -t "Add JWT authentication to the Flask app" -f app.py

This starts the full 8-phase team session:

PLANNING -- Planner decomposes the task into subtasks
DISCUSSION -- All advisors share their perspective (parallel)
CODING -- Coder writes the actual production code
REVIEW -- Architect + Debugger + Security review code (parallel)
DISSENT -- Observers + Coordinator stress-test consensus (mandatory)
APPROVAL -- Human accepts, rejects, or edits the proposed changes
TESTING -- Auto-run tools (pytest, ruff, mypy) on the changes
DONE -- Changes applied or session complete

3. Cowork Mode (watch your project)

crosscheck team -t "Review and improve" --project . --auto-apply --watch

The team watches your project directory for changes and provides real-time feedback as you code in your editor.

4. Open the Group Chat UI (web browser)

crosscheck chat --port 8080

This opens a web-based group chat at http://localhost:8080 where you can:

See all 9 agents talking in real-time
Send messages to the whole team
Use @coder, @debugger, @security to target specific agents
Watch the phase progression in the sidebar
Export the full transcript as markdown

5. Code Review (simpler, review-only mode)

crosscheck review mycode.py

Multiple AI models review your code from different angles (bugs, security, architecture, performance). Two supervisors vote independently using conservative reconciliation.

6. Observer Mode (watch external sessions)

# Watch a folder for changes
crosscheck observe watch ./src

# Paste code directly
crosscheck observe paste --code "def login(user, pw): ..."

How the Team Works

User: "Add input validation to the API endpoints"
         |
         v
  [PLANNER / CEO] -- Breaks into subtasks, assigns to team
         |
         v
  [DISCUSSION] -- All advisors share perspective (parallel)
    - Architect: "Use Pydantic models at the boundary"
    - Debugger: "Watch for empty string vs None edge cases"
    - Security: "Validate against injection patterns"
    - Analyst: "Keep validation fast, don't add latency"
         |
         v
  [CODER] -- Writes actual production code
    - Follows team advice
    - Outputs complete files with filenames
         |
         v
  [REVIEW] -- Architect + Debugger + Security review (parallel)
         |
         v
  [DISSENT] -- Observers + Coordinator stress-test consensus
    - Observer 1: Finds edge cases team missed
    - Observer 2: Different perspective (diverse model)
    - Coordinator: Final safety gate
         |
         v
  [APPROVAL] -- Human reviews and decides
    - Accept -> apply changes
    - Reject -> back to CODING
    - Edit -> modify before applying
         |
         v
  [TESTING] -- Auto-run pytest, ruff, mypy on changes
    - Pass -> DONE
    - Fail -> auto-rollback, back to CODING

Default Team (all FREE)

Role	Model	Context
Coordinator	Qwen 3.6 Plus (Free)	1M
Planner (CEO)	Nemotron 3 Super 120B (Free)	262K
Architect (CTO)	Qwen 3.6 Plus (Free)	1M
Coder	Qwen 3 Coder (Free)	262K
Debugger	Nemotron 3 Super 120B (Free)	262K
Security Lead	Qwen 3 Next 80B (Free)	131K
Flow Analyst	Trinity Large (Free)	131K
Observer 1	Nemotron 3 Super 120B (Free)	262K
Observer 2	Hermes 3 Llama 405B (Free)	131K

Override any role with premium models:

crosscheck team -t "task" --models "coder:anthropic/claude-sonnet-4.6,architect:openai/gpt-5"

InsAIts Security Integration

crosscheck-ai integrates with InsAIts to monitor every AI-to-AI message for:

Prompt Injection -- "IGNORE ALL INSTRUCTIONS" patterns -> BLOCKED (critical)
Credential Leaks -- API keys, passwords in agent output -> DETECTED (high)
Jargon Drift -- Agents using undefined acronyms/shorthand -> DETECTED (medium)
Hallucination -- Nonsensical or ungrounded claims -> DETECTED (medium)

Install with monitoring:

pip install "crosscheck-ai[monitor]"

InsAIts activates automatically when installed -- no configuration needed.

@Mentions

In both the Group Chat UI and the terminal, use @ to target specific agents:

@coder fix the null check on line 45
@security scan the auth middleware for vulnerabilities
@debugger @analyst both check the database query performance
@architect should we use microservices or monolith?
@coordinator is the team consensus solid?

Aliases work too:

@boss, @ceo, @lead -> Planner
@cto, @designer -> Architect
@dev, @builder -> Coder
@debug, @qa, @tester -> Debugger
@sec, @guard -> Security
@flow, @perf -> Analyst
@coord, @safety, @gate -> Coordinator
@obs1, @obs2 -> Observers

Model names also work: @claude, @gpt, @gemini, @grok, @deepseek, @qwen

All CLI Commands

# AI Dev Team
crosscheck team -t "task" [-f file.py] [--rounds 3] [--output terminal|json|transcript]
crosscheck team -t "task" --project . --auto-apply --watch     # Cowork mode
crosscheck chat [--port 8080] [--no-open]                      # Group Chat UI

# Code Review
crosscheck review FILE [--mode fast|balanced|quality] [--max-rounds 3]
crosscheck review FILE --show-diff [--show-diff-mode side-by-side]
crosscheck review FILE --repo-context

# Observer
crosscheck observe paste --code "..."
crosscheck observe watch ./src [--plan "expected behavior"]

# PR Bot
crosscheck pr review --repo owner/repo --pr 42
crosscheck pr webhook --port 8080

# Tools & Sandbox
crosscheck sandbox [PATH] [--tools ruff,pytest] [--list]

# Dashboard
crosscheck dashboard [--port 8080]

# Cache
crosscheck cache stats
crosscheck cache clear [--expired-only]

# Policy
crosscheck policy list
crosscheck policy validate policy.yaml

# Local Models (Ollama)
crosscheck local list-models [--url http://localhost:11434]
crosscheck local review FILE [--supervisor model] [--analyzer model]

# Configuration
crosscheck init            # Create crosscheck.toml
crosscheck profiles        # List profiles
crosscheck models list     # Show all 36 models
crosscheck types           # List review types

Configuration

Environment Variables

CROSSCHECK_API_KEY       # Required -- OpenRouter API key
OPENROUTER_API_KEY       # Alias for CROSSCHECK_API_KEY
INSAITS_API_KEY          # Optional -- InsAIts monitoring key
CROSSCHECK_MODE          # fast | balanced | quality
CROSSCHECK_MAX_ROUNDS    # Max review/coding rounds (default: 3)
CROSSCHECK_OUTPUT        # terminal | json | markdown
CROSSCHECK_MONITOR       # true | false (InsAIts monitoring)

Config File (crosscheck.toml)

crosscheck init  # Creates crosscheck.toml with example config

[default]
api_key = "sk-or-v1-..."
mode = "balanced"
max_rounds = 3
output = "terminal"

[profiles.fast]
mode = "fast"
max_rounds = 1
supervisors = ["anthropic/claude-haiku-4.5"]

[profiles.quality]
mode = "quality"
max_rounds = 5
supervisors = ["anthropic/claude-opus-4.6", "openai/gpt-5"]

Architecture

crosscheck/
  cli.py + cli_extensions.py    CLI (14+ commands)
  client.py                     OpenRouter async HTTP client
  config.py                     TOML + env config, profiles
  core.py                       MultiAgentSession orchestrator
  models.py                     36-model registry
  monitor.py                    InsAIts integration (real-time AI security)
  observer.py                   Observer: paste + watch
  prompts.py                    Agent system prompts
  reporter.py                   Terminal/JSON/Markdown output
  cache.py                      Smart caching + AST delta
  context_builder.py            Multi-file repo context
  dashboard.py                  FastAPI dashboard + SQLite
  chat_server.py                WebSocket Group Chat server
  chat_ui/                      Vanilla HTML/JS/CSS chat UI
  diff.py                       Unified + side-by-side diffs
  local.py                      Ollama/LM Studio support
  policy.py                     Company policy agents
  pr_bot.py                     GitHub/GitLab PR bot
  sandbox.py                    Tool sandbox (pytest, ruff, mypy...)
  streaming.py                  Streaming events + Rich Live UI
  agents/
    supervisor.py               Dual-voting supervisor engine
    analyzer.py                 Parallel analyzer pool
    coder.py                    Any-model coder
  team/
    roles.py                    TeamRole enum, RoleSpec, DEFAULT_TEAM (9 agents)
    chat.py                     TeamMessage, ChatHistory, CodeBlock
    command_parser.py           @mention routing + aliases
    language.py                 20+ language auto-detection
    session.py                  TeamSession orchestrator (8 phases)
    workspace.py                ProjectWorkspace (safe file ops + rollback)

Running Tests

# All tests (283 tests, all mocked, no API key needed)
pytest tests/ -v --tb=auto

# Just the team tests (108 tests)
pytest tests/test_team.py -v

# With coverage
pytest tests/ --cov=crosscheck --cov-report=term-missing

PyPI Package

The package is published on PyPI as crosscheck-ai:

pip install crosscheck-ai          # Basic
pip install "crosscheck-ai[chat]"  # With Group Chat UI
pip install "crosscheck-ai[all]"   # Everything (InsAIts + chat + policy)

To build from source:

python -m build
twine check dist/*
twine upload dist/*  # Requires PyPI credentials

License

This is proprietary software. Unauthorized copying, distribution, or modification is strictly prohibited.

Contact: info@yuyai.pro

Name		Name	Last commit message	Last commit date
crosscheck		crosscheck
tests		tests
vscode-extension		vscode-extension
.env.example		.env.example
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
pyproject.toml		pyproject.toml

Navigation Menu

Repository navigation

About

Repository files navigation

crosscheck-ai

What is crosscheck-ai?

Features

Install

Basic install (CLI + team + review)

With Group Chat UI (recommended)

Full install (all features including InsAIts monitoring)

For development

Quick Start

1. Set your API key

2. Launch the AI Dev Team (terminal)

3. Cowork Mode (watch your project)

4. Open the Group Chat UI (web browser)

5. Code Review (simpler, review-only mode)

6. Observer Mode (watch external sessions)

How the Team Works

Default Team (all FREE)

InsAIts Security Integration

@Mentions

All CLI Commands

Configuration

Environment Variables

Config File (crosscheck.toml)

Architecture

Running Tests

PyPI Package

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

crosscheck-ai

What is crosscheck-ai?

Features

Install

Basic install (CLI + team + review)

With Group Chat UI (recommended)

Full install (all features including InsAIts monitoring)

For development

Quick Start

1. Set your API key

2. Launch the AI Dev Team (terminal)

3. Cowork Mode (watch your project)

4. Open the Group Chat UI (web browser)

5. Code Review (simpler, review-only mode)

6. Observer Mode (watch external sessions)

How the Team Works

Default Team (all FREE)

InsAIts Security Integration

@Mentions

All CLI Commands

Configuration

Environment Variables

Config File (crosscheck.toml)

Architecture

Running Tests

PyPI Package

License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages