98–99% KV-cache hit rate. Native code graph. Quad-protocol native: OpenAI · Responses · Anthropic · Gemini. Extreme self-healing tool chain.

JeikCode: Ultra-Fast, Autonomous AI Coding Agent (Rust-Driven)

98–99% KV-Cache · Native Code Graph · Quad-Protocol Native · Extreme Tool Self-Healing

Next-generation Agentic AI coding assistant purpose-built for complex, large-scale production codebases.

English (Default) · 简体中文 (Chinese)

What is JeikCode · Architecture Comparison · CodeExplore Indexing · Highlights · Installation · Commands · Knowledge Packs

🌐 Official GitHub Repository: https://github.com/jeikl/JeikCode

1. What is JeikCode?

JeikCode is a zero-bloat, ultra-fast autonomous AI coding agent natively built with Rust, specifically engineered to eliminate context bloat, eliminate blind grepping, and master large-scale complex software architectures.

Four defining weapons sit in front of everything else:

1. 🔥 Terrifying KV-cache hit rate:
   • 98–99% on almost every conversation, because prefixes stay byte-exact across turns.
   • Strict Append-Only Prefix Discipline: user queries are wrapped only at the tail (user-wrap.md). Paired with sacred_floor memory compaction protection, it eliminates KV-cache thrashing and slashes LLM inference cost.
   • Core prompts live in init.yaml, rules.yaml, and user-wrap.md — hot-reload on save, no restart, no rebuild.
2. 🗺️ Native code graph (CodeExplore):
   • Built-in weighted AST syntax graphs + bilingual natural-language / docstring semantic alignment (Cilin-weighted).
   • Query by business logic (e.g. "Find refund callback error handling") instead of blind regex or symbol-only LSP lookups — 60%–70% faster retrieval, 90%+ location accuracy.
3. 🔌 Quad-protocol native:
   • One agent speaks OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and Google Gemini natively.
   • Swap providers without rewriting the agent loop.
4. 🩸 Extreme self-healing tool chain:
   • Five-stage auto-heal for JSON, types, Windows paths, truncated payloads, and edit_file mismatches — the model is treated as a fallible child, not a perfect caller.
   • Same-file multi-hunk edits stay atomic (WAR topology + all-or-nothing write). Cross-turn VersionRing 3-way rebase recovers stale old_string without hanging the runtime.

2. Functional & Architectural Comparison

The following matrix objectively evaluates JeikCode, OpenAI Codex, Claude Code, OpenCode, and Grok Build based on real-world engineering benchmarks:

1. Feature & Mechanism Comparison Matrix

2. Programming Language AST Support Matrix

3. Native CodeExplore & repo_map Deep Retrieval

While CodeGraph introduced valuable symbol-indexing concepts, real-world engineering surfaced a critical bottleneck: it only understands rigid code symbols and has zero semantic understanding of natural language. When a developer asks "Where is the refund callback verified?", pure symbol search is blind.

JeikCode developed a fully autonomous CodeExplore and repo_map engine:

1. Weighted AST Vectors + Semantic Alignment:
   • Parses code structures (AST symbols, call graphs, type definitions, trait bounds).
   • Extracts bilingual comments and docstrings.
   • Aligns code logic with natural language business semantics using semantic embedding and Cilin lexical weighting.
2. Weighted Relevance Ranking:
   • Scores and directly pins the most relevant implementation blocks at the top of the agent context, eliminating trial-and-error grep loops.
3. Budget-Aware Low-Relevance File Recommendations:
   • Never dumps entire secondary files into context. Instead, condenses them into minimal token-budget paths and structural summaries, preserving maximum context space for active development.
4. Benchmarked Metrics:
   • 🚀 60% - 70% reduction in search latency: Pinpoints exact targets in a single round;
   • 🎯 90%+ retrieval accuracy: Handles bilingual and abstract business queries seamlessly.

4. Core Architectural Highlights

1. Strict Append-Only Cache Stability & Lean Prompts

• Byte-Exact Immutability: System identity, MEMORY, SKILLS, and knowledge rules are packed compactly at the conversation header.
• user-wrap.md Dynamic Tail Wrapping: The {{input}} template dynamically wraps only the active turn's user message. Modifying the template hot-reloads in milliseconds without invalidating cached prefix tokens.
• sacred_floor Compaction Guard: During /compact context summarization, critical rules and memory entries are anchored below the floor and are never dropped.

2. 5-Stage Tool Recovery, Self-Healing & Loop Guard

• 5-Stage Tool Self-Healing: Direct parsing → Lenient JSON repair (trailing commas, unquoted keys, markdown code fences stripping) → edit_file regex rescue → Schema string decoding → Key-value fallback.
• Windows Path Backslash Sanitization: Automatically rescues unescaped single backslashes in paths like D:\project\src before serde deserialization.
• Automatic Type Coercion: Automatically coerces "count":"5" to 5 and "verbose":"true" to true.
• 64KB Large Output Folding: Massive outputs fold into artifacts with head/tail previews; agent queries slices via fetch_output as needed.
• 3-Attempt Loop Guard Circuit Breaker: Tripped when identical tool calls fail 3 times sequentially, halting runaway loops and forcing alternative strategies.
• Dual-Tier Timeout Architecture: 1800s hard process lifetime with process tree reclamation; 180s progress-aware idle budget (reset on notifications/progress).
• Independent First-Token Watchdog: 60s × 3 independent retry watchdog for reasoning models (DeepSeek-R1, Grok 3), preventing silent stream hangs.

3. Native Multi-Protocol Support & 4-Gear Reasoning

• Native support for four primary model protocols: OpenAI Responses (/v1/responses), OpenAI Chat Completions, Anthropic Messages, and Google Gemini Native Protocol.
• Switch reasoning effort on the fly with /effort (low / medium / high / xhigh / off).
• Model credentials and parameters are fully decoupled; /modeladd queries upstream /models dynamically.

4. Modern Interactive WebUI & Multi-Instance Headless

• WebUI Console: Launch interactive browser control via /webui or jeikcode webui (featuring KaTeX LaTeX rendering and real-time token cost overlays).
• Streaming Multipart Attachments: Drag-and-drop multimodal attachments with upload progress. Non-image files stream into .jeikcode_store and remain indexed for tool queries.
• Multi-Instance Headless Serve:

  jeikcode serve --host 0.0.0.0 --port 4096 --token sk-my-secret
  jeikcode attach http://192.168.1.100:4096 --token sk-my-secret
  

5. Installation & Quick Start

1. One-Line Script Installation (Recommended)

Prebuilt binaries are available on GitHub Releases:

# Linux / macOS Installation
curl -fsSL https://raw.githubusercontent.com/jeikl/JeikCode/main/scripts/install.sh | bash

# Windows PowerShell Installation
irm https://raw.githubusercontent.com/jeikl/JeikCode/main/scripts/install.ps1 | iex

2. Upgrades & Custom Update Channel

JeikCode natively supports seamless in-place self-upgrades (update and upgrade are completely identical):

# Check and upgrade in-place to latest release
jeikcode update

# Automatic non-interactive upgrade (auto-sync prompts/teaches, protect custom config)
jeikcode update -y

# Configure and remember custom/mirror update source (persists to ~/.jeikcode/config.toml)
jeikcode update set https://github.com/jeikl/JeikCode

# Inspect current update source and resolved endpoints
jeikcode update get

# Roll back to the previous version
jeikcode update rollback

Or re-run the one-line installer at any time to overwrite with the latest release:

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/jeikl/JeikCode/main/scripts/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/jeikl/JeikCode/main/scripts/install.ps1 | iex

3. Build from Source

Prerequisites: Rust 1.88+ (rustup.rs):

git clone https://github.com/jeikl/JeikCode.git
cd JeikCode

cargo install --path crates/jeikcode-cli --bin jeikcode --locked
jeikcode --version

4. Configuration & Launch

Launch directly from any project directory:

cd /path/to/your/project
jeikcode

Configuration file is located at ~/.jeikcode/config.toml:

default_provider = "deepseek"

[provider_accounts.deepseek]
api_key  = "sk-xxxxxxxxxxxxxxxxxxxxxxxx"
base_url = "https://api.deepseek.com/v1"

[models.deepseek-chat]
provider = "deepseek"
model    = "deepseek-chat"
protocol = "chat_completions"

[models.deepseek-reasoner]
provider         = "deepseek"
model            = "deepseek-reasoner"
protocol         = "chat_completions"
reasoning_effort = "high"

Common CLI commands:

# Launch in a specific project directory
jeikcode -C /path/to/project

# Launch with a specific model
jeikcode --model deepseek-reasoner

# Headless mode for automated scripts / CI/CD
jeikcode -p "Investigate and fix OAuth callback 404 error"

# Resume previous conversation
jeikcode -c

6. Keybindings & Commands

1. Terminal Shortcuts

2. Slash Commands

7. Multi-Project Knowledge Packs

JeikCode supports multi-tier project rules that hold strict execution precedence over System default prompts:

Changelog

v7.1.34 (2026-09-29): Desktop app suppresses the Windows background console window; ships official high-res JeikCode logo icons; fixes input footer status bar text collision in narrow/un-maximized windows via container responsive layout.

Full notes: CHANGELOG.md.

License

This project is licensed under the JeikCode Non-Commercial & Attribution License (Adapted from CC BY-NC 4.0). See the full LICENSE for details:

1. Non-Commercial Restriction: The Software and its derivative works MAY NOT be used, in whole or in part, for any commercial purpose, commercial product integration, monetized services, SaaS hosting, or commercial redistribution without prior written authorization from the author.
2. Mandatory Attribution: Any public deployment, distribution, fork, or derivative work MUST prominently retain and display the original copyright notice, the project name (JeikCode), author credit, and the official repository URL: https://github.com/jeikl/JeikCode.
3. Commercial Licensing: For commercial licensing, enterprise integration, or partnership inquiries, please contact the author via the official repository.