"Finally, an AI coding assistant that doesn't just guess — it proves."

Clyde brings Formation's breakthrough Stratus X1 reasoning model to your terminal.

Traditional AI coding assistants are amazing at understanding code but terrible at verification. They guess. They approximate. They fail at counting. Clyde fixes this by combining Claude's language mastery with Stratus X1's mathematical rigor.

Built on Anthropic's Claude Code, Clyde supercharges your coding workflow with Stratus X1 — Formation's T-JEPA (Temporal Joint-Embedding Predictive Architecture) reasoning model that delivers transformational improvements where LLMs traditionally fail:

• Perfect counting: Count 'r' in 'strawberry' → 3 ✅ (vs Claude's 2 ❌)
• Flawless logic: Multi-step reasoning with mathematical proof
• Exact verification: Prime number testing, regex validation, code analysis

• Prime number verification: <2 seconds vs minutes of LLM chain-of-thought
• Token counting: Instant vs unreliable approximations
• Code complexity analysis: Real-time vs expensive multi-pass processing

• Analyze 50k+ line files without hitting context limits
• Maintain full codebase understanding across 100+ files
• T-JEPA latent space compression preserves semantic meaning

• Multi-step task decomposition with lookahead
• Dependency graph inference from code structure
• Intelligent task prioritization based on complexity

Clyde is the only AI coding assistant that combines Claude's language understanding with mathematically rigorous reasoning.

💡 Clyde is production-ready. Drop-in replacement for Claude Code with zero configuration needed.

Original Claude Code documentation: code.claude.com

Run the installer:

cd /Users/andrewhathaway/code/formation/clyde-code
./install.sh

Then use clyde anywhere:

cd ~/your-project
clyde

Clyde works exactly like claude but with Stratus X1 reasoning built-in.

See INSTALL.md for full installation guide.

Every. Single. Time. Correct.

Large Language Models (LLMs) are incredible at language understanding but fundamentally broken at:

• Counting tokens, characters, or occurrences
• Mathematical reasoning without expensive chain-of-thought
• Understanding massive codebases without losing context
• True multi-step planning with verification

Stratus X1 solves these problems at the architecture level.

Formation's Stratus X1 uses T-JEPA (Temporal Joint-Embedding Predictive Architecture) — a fundamentally different approach than traditional transformers:

• Latent Space Reasoning: Compresses information into abstract representations that preserve exact semantics
• Predictive Architecture: Forecasts future states instead of just predicting next tokens
• Mathematical Rigor: Provides provable guarantees on counting, verification, and logic
• Efficient Scaling: O(log n) complexity on many tasks where LLMs are O(n²)

This isn't incremental improvement — it's a paradigm shift.

# Count occurrences in a massive log file
> How many times does "ERROR" appear in server.log?
Stratus X1: 1,247 (verified in 0.8s)
Claude: "approximately 1,200-1,300" (best guess after 30s)

# Verify complex regex patterns
> Does this regex match all valid email addresses?
Stratus X1: No - fails on unicode domains (proof attached)
Claude: "Yes, this should work for most cases"

# Analyze huge codebases
> Find all authentication flows in this 200k line project
Stratus X1: 7 flows identified with call graphs (3.2s, 2k tokens used)
Claude: Context limit exceeded after 3 files

See the Stratus Integration Guide for technical details and benchmarks.

The magic happens in the stratus-reasoning plugin, which seamlessly integrates Stratus X1 into Claude Code's agentic workflow:

• Automatic Detection: Clyde automatically routes counting, verification, and reasoning tasks to Stratus X1
• Transparent Integration: Works exactly like Claude Code — no special commands needed
• Hybrid Reasoning: Combines Claude's language understanding with Stratus X1's mathematical rigor
• Efficient Fallback: Uses Claude for language tasks, Stratus X1 for verification tasks

Technical capabilities:

• 🔢 Counting: Characters, tokens, occurrences, patterns
• ✅ Verification: Primes, regex, logic, mathematical proofs
• 📦 Compression: 10-50x token reduction for massive files
• 🧭 Planning: Multi-step task decomposition with dependency graphs

See the Stratus Reasoning Plugin Guide for API details and benchmarks.

Clyde inherits all Claude Code plugins and adds several new ones:

• clyde-core - Core Clyde functionality and Stratus integration
• All original Claude Code plugins (see plugins directory)

Clyde is perfect for developers who:

• 🏗️ Work with massive codebases (50k+ lines) that exceed LLM context limits
• 🔍 Need exact verification instead of "probably correct" answers
• ⚡ Want instant results on counting, pattern matching, and analysis tasks
• 🧪 Write tests and validators that require mathematical correctness
• 📊 Deal with large datasets and need accurate counting/parsing
• 🚀 Value speed and accuracy over approximate results

Clyde is NOT for you if:

• ❌ You only need basic code completion (use GitHub Copilot)
• ❌ You prefer web UIs over terminal tools
• ❌ You don't care about exact correctness for counting/verification tasks

Try these tasks that showcase Stratus X1's capabilities:

# Perfect counting (LLMs typically fail this)
> Count the letter 'r' in "strawberry"

# Fast prime verification with proof
> Is 179424673 prime?

# Massive file analysis with compression
> Analyze all authentication flows in this 50k line codebase

# Complex regex verification with counterexamples
> Verify this regex matches all valid IPv6 addresses

# Exact token counting for cost estimation
> How many tokens are in these 50 API response files?

We welcome your feedback. File a GitHub issue to report problems or suggest improvements.

For issues with the underlying Claude Code functionality, see the original repository.

Formation is pioneering the next generation of AI systems that go beyond language models.

Stratus X1 represents years of research into Joint-Embedding Predictive Architectures — an approach championed by Yann LeCun and now realized at production scale by Formation:

• Architecture: T-JEPA (Temporal Joint-Embedding Predictive Architecture)
• Innovation: Learns abstract representations of temporal sequences without next-token prediction
• Capability: Achieves superhuman performance on verification, counting, and reasoning tasks
• Integration: Seamlessly augments LLMs like Claude with mathematical rigor

Clyde demonstrates what's possible when you combine language understanding (Claude) with verified reasoning (Stratus X1).

• Clyde: Developed by Formation on top of Claude Code
• Stratus X1: Formation's breakthrough T-JEPA reasoning model
• Claude Code: Anthropic's excellent AI coding assistant (github.com/anthropics/claude-code)
• License: Same as Claude Code (see LICENSE.md)

Want to integrate Stratus X1 into your products? Contact Formation →

Formation is actively developing new Stratus X1 capabilities:

• 🔬 Code Analysis: Static analysis with mathematical guarantees (Q2 2026)
• 🎯 Bug Detection: Formal verification of correctness properties (Q2 2026)
• 🚀 Performance Prediction: Actual complexity analysis vs asymptotic estimates (Q3 2026)
• 🧠 Multi-Modal Reasoning: Extend T-JEPA to understand diagrams and architecture (Q3 2026)
• 🌐 Distributed Reasoning: Parallel verification across compute clusters (Q4 2026)

This is just the beginning. T-JEPA architectures unlock capabilities impossible with traditional LLMs.

• ⭐ Star this repo to stay updated on new releases
• 🐛 Report issues at github.com/hathbanger/clyde-code/issues
• 💬 Discuss ideas in GitHub Discussions
• 🔗 Share your success stories — tag @FormationAI on Twitter

Want to contribute? We welcome PRs that improve Clyde or extend Stratus X1 integration!

Note: As a fork of Claude Code, Clyde inherits Claude Code's data policies for the underlying infrastructure.

When you use Clyde:

• Anthropic (Claude Code): Data is handled according to Claude Code's data usage policies
• Formation (Stratus X1): Queries sent to Stratus X1 are processed according to Formation's privacy policy

For full details on Claude Code data handling, please review Anthropic's Commercial Terms of Service and Privacy Policy.

For Stratus X1 data handling, contact Formation at team@formation.cloud.