ABCoder, an AI-oriented Code-processing Framework, is designed to enhance and extend the coding context for Large-Language-Model (LLM), finally boosting the development of AI-assisted-programming applications.

• Universal Abstract-Syntax-Tree (UniAST), a language-independent, AI-friendly specification of code information, providing a boundless, flexible and structural coding context for both AI and humans.

• General Parser, parses arbitrary-language codes to UniAST.

• General Writer transforms UniAST back to code.

• Code-Retrieval-Augmented-Generation (Code-RAG), provides a set of MCP tools to help the LLM understand code repositories precisely and locally. And it can support both in-workspace and out-of-workspace third-party libraries simultaneously -- I guess you are thinking about DeepWiki and context7, but ABCoder is more reliable and confidential -- no need to wait for their services to be done, and no worry about your codes will be uploaded!

Based on these features, developers can easily implement or enhance their AI-assisted programming applications, such as reviewing, optimizing, translating, etc.

see UniAST Specification

ABCoder provides deep integration with Claude Code through the AST-Driven Coding workflow, enabling hallucination-free code analysis and precise execution. Check Claude Code Specification for more details.

Use the init-spec command to automatically configure Claude Code integration for your project:

# Install ABCoder
go install github.com/cloudwego/abcoder@latest

# Run init-spec in your project directory (optional: specify target path)
cd /path/to/your/project
abcoder init-spec

The init-spec command will:

1. Copy .claude directory to your project root
2. Configure MCP servers in ~/.claude.json:
   • abcoder: for code analysis using AST
   • sequential-thinking: for complex problem decomposition
3. Replace all {{CLAUDE_HOME_PATH}} placeholders with actual project paths

Once setup, you can start coding with Claude Code:

1. Start Claude Code in your project directory
2. Use slash common /abcoder:schedule <problem_desc> to address your feature/requirement/issue, and ABCoder will help you analyze the codebase and design a technical solution.
3. Once all questions are set, use slash common /abcoder:task <task_name> to create a coding task(specification)
4. Recheck the task using /abcoder:recheck <task_name> before real implementation
5. Begin coding! Claude Code will process the task step by step according to the specification, leveraging the power of AST-driven analysis.

.claude/hooks provide a 4-layer analysis chain from repository to node details:

list_repos → get_repo_structure → get_package_structure → get_file_structure → get_ast_node
     │              │                      │                       │                    │
     └── repo_name  └── mod/pkg list       └── file list           └── node list        └── dependencies/references

.claude/commands provide three custom slash commands to streamline development:

User Request
    │
    ▼
/abcoder:schedule ──────────→ Design Solution (ABCoder Analysis)
    │                          │
    ▼                          ▼
/abcoder:task ─────────→ CODE_TASK (with Technical Specs, including accurate `get_ast_node` call args)
    │                          │
    ▼                          ▼
/abcoder:recheck ────→ Verify Solution (ABCoder Validation. After `/abcoder:task` Claude Code will tell you what the external dependencies CODE_TASK contains, use `/abcoder:recheck` to analyze external ast_node and technical detail with ABCoder)
    │                          │
    ▼                          ▼
Start coding(sub-agent) ─────────→ Execute Implementation

• Claude Code CLI
• ABCoder MCP server (provides mcp__abcoder tools)
• Sequential-thinking MCP server (provides mcp__sequential_thinking tools, automatically configured by init-spec)

For detailed configuration, see claude-code-spec.md

Watch the demo video here

The Skill interface provides native Claude Code integration without MCP protocol overhead. It uses Claude Code's built-in skill system for a more streamlined workflow.

The Skill is automatically configured when you run abcoder init-spec. The skill definitions are located in internal/cmd/assets/.claude/skills/.

# List all repositories
abcoder cli list_repos

# Get repository file tree
abcoder cli tree_repo 'repo_name'

# Get file structure
abcoder cli get_file_structure 'repo_name' 'path/to/file.go'

# Get symbol details
abcoder cli get_file_symbol 'repo_name' 'path/to/file.go' 'SymbolName'

# Search symbols
abcoder cli search_symbol 'repo_name' 'Pattern*'

For Claude Code integration, the skill tools are invoked directly via slash commands like /abcoder:schedule.

1. Install ABCoder:

   go install github.com/cloudwego/abcoder@latest

2. Use ABCoder to parse a repository to UniAST (JSON)

   abcoder parse {language} {repo-path} -o xxx.json

   ABCoder will try to install any dependency automatically. In case of failure (or if you want to customize installation), refer to the docs.

   For example, to parse a Go repository:

   git clone https://github.com/cloudwego/localsession.git localsession
   abcoder parse go localsession -o /abcoder-asts/localsession.json

3. Integrate ABCoder's MCP tools into your AI agent.

   {
       "mcpServers": {
           "abcoder": {
               "command": "abcoder",
               "args": [
                   "mcp",
                   "{the-AST-directory}"
               ]
           }
       }
   }

4. Enjoy it!

   Try to click and watch the video below:

   

• You can add more repo ASTs into the AST directory without restarting abcoder MCP server.

• Try to use the recommended prompt and combine planning/memory tools like sequential-thinking in your AI agent.

ABCoder provides two integration methods with Claude Code:

The Skill interface is the recommended approach for Claude Code users, providing a more streamlined workflow:

• Auto-detect current repo: list_repos automatically detects repos that match current working directory
• Memory efficient: Uses Sonic for lazy JSON parsing, only loads needed data
• Pipeline friendly: Output can be piped to rg for filtering or jq for extraction
• Regex search: search_symbol supports regex patterns to precisely locate symbols

Pipeline Examples:

# Filter current repo(s) only
abcoder cli list_repos | jq '.current_repo'

# Search with regex
abcoder cli search_symbol myrepo "^Get.*User$"

# Filter related file
abcoder cli tree_repo myrepo | rg 'related-file'

# Filter dependencies only
abcoder cli get_file_symbol myrepo src/main.go MyFunc | jq '.node.dependencies'

For detailed usage, see Skill Definitions.

You can also use ABCoder as a command-line Agent like:

export API_TYPE='{openai|ollama|ark|claude}' 
export API_KEY='{your-api-key}' 
export MODEL_NAME='{model-endpoint}' 
abcoder agent {the-AST-directory}

For example:

$ API_TYPE='ark' API_KEY='xxx' MODEL_NAME='zzz' abcoder agent ./testdata/asts

Hello! I'm ABCoder, your coding assistant. What can I do for you today?

$ What does the repo 'localsession' do?

The `localsession` repository appears to be a Go module (`github.com/cloudwego/localsession`) that provides functionality related to managing local sessions. Here's a breakdown of its structure and purpose:
...
If you'd like to explore specific functionalities or code details, let me know, and I can dive deeper into the relevant files or nodes. For example:
- What does `session.go` or `manager.go` implement?
- How is the backup functionality used?

$ exit

• NOTICE: This feature is Work-In-Progress. It only supports code analysis at present.

ABCoder currently supports the following languages:

We encourage developers to contribute and make this tool more powerful. If you are interested in contributing to ABCoder project, kindly check out our guide:

• Parser Extension

Note: This is a dynamic README and is subject to changes as the project evolves.

• How to become a member: COMMUNITY MEMBERSHIP
• Issues: Issues
• Lark: Scan the QR code below with Register Feishu to join our CloudWeGo/abcoder user group.

Thank you for your contribution to ABCoder!

This project is licensed under the Apache-2.0 License.