README.md

TaskWing

The local-first knowledge layer for AI development.

Website · Tutorial · Vision · Install

---

Your AI tools start every session from zero -- and every session, your code context flows through someone else's cloud.

TaskWing takes the opposite approach. One command extracts your architecture into a local knowledge base on your machine. No cloud. No account. Every AI session after that just knows -- without your knowledge base leaving your infrastructure.

Without TaskWing              With TaskWing
─────────────────             ─────────────
8-12 file reads               1 taskwing ask
~25,000 tokens                ~1,500 tokens
2-3 minutes                   42 seconds
No architectural context       170+ knowledge nodes

Install

brew install josephgoksu/tap/taskwing

No signup. No account. Works offline. Everything stays local in SQLite.

Alternative: install via curl

curl -fsSL https://taskwing.app/install.sh | sh

Quick Start

# 1. Declare a TaskWing project + generate AI tool integration files
cd your-project
taskwing init                  # writes .taskwing.yaml + .claude/commands/taskwing/

# 2. Extract your architecture (one-time)
taskwing learn
# -> 22 decisions, 12 patterns, 9 constraints extracted

# 3. Plan and execute with your AI assistant
/taskwing:plan       # Create a plan
/taskwing:next       # Get next task with full context
# ...work...
/taskwing:done       # Mark complete, advance to next

That's it. Your AI assistant now has local architectural context across every session.

Slash commands are pure prompts that drive the taskwing CLI directly - no MCP server, no daemon, no per-AI registration. They work in any AI client that can run shell commands.

Private by Architecture

TaskWing keeps your knowledge base on your machine. No cloud database, no account, no sync.

  YOUR MACHINE                          EXTERNAL
  ─────────────────────────────────     ─────────────────────────
                                        ┌───────────────────────┐
  ┌──────────────┐   code context       │ LLM Provider          │
  │ Your codebase ├────────────────────>│ (OpenAI, Anthropic,   │
  └──────────────┘   (bootstrap only)   │  Google, Bedrock)     │
         │                              └───────────┬───────────┘
         │                                          │ findings
         v                                          │
  ┌──────────────────────┐  <───────────────────────┘
  │ .taskwing/memory.db  │
  │ Local SQLite         │  Your knowledge base.
  │ Never uploaded.      │  Never leaves your machine.
  └──────────┬───────────┘
             │ taskwing CLI (local subprocess)
             v
  ┌──────────────────────┐              ┌───────────────────────┐
  │ AI Tool              │  may send    │ Tool's own cloud      │
  │ (Claude, Cursor,     ├─────────────>│ (per their privacy    │
  │  Copilot, Gemini)    │  to their    │  policy)              │
  └──────────────────────┘  servers     └───────────────────────┘


  FULL AIR-GAP (everything stays left of the line):

  ┌──────────────┐        ┌─────────┐        ┌──────────────┐
  │ Your codebase ├──────>│ Ollama  ├──────>│ .taskwing/   │
  └──────────────┘        │ (local) │        │ memory.db    │
                          └─────────┘        └──────┬───────┘
                                                    │ taskwing CLI
                                                    v
                                             ┌──────────────┐
                                             │ Local AI tool │
                                             └──────────────┘
                                             Zero network calls.

What TaskWing controls: Your knowledge base is stored and queried locally. AI tools invoke the taskwing CLI as a local subprocess - no network calls.

What your AI tool controls: Cloud-based tools (Claude, Cursor, Copilot) may send conversations to their own servers. Check their privacy settings (e.g., Cursor's Privacy Mode, Copilot's data retention policies).

Full air-gap: Use Ollama for bootstrap + a local AI tool. Nothing leaves your machine.

Works With

Supported Models

Brand names and logos are trademarks of their respective owners; usage here indicates compatibility, not endorsement.

What It Does

Capability	Description
Local knowledge	Extracts decisions, patterns, and constraints into local SQLite
Plan to tasks	Turns a plan into decomposed tasks with architecture context
AI-driven lifecycle	Task execution -- next, start, complete, verify
Code analysis	Symbol search, call graphs, impact analysis, simplification
Root cause first	AI-powered diagnosis before proposing fixes
Works everywhere	Slash commands invoke the taskwing CLI directly - works in any AI tool that runs shell commands

Slash Commands

Use these from your AI assistant once connected:

Command	When to use
/taskwing:plan	Clarify a goal and build an approved execution plan
/taskwing:next	Start the next approved task with full context
/taskwing:done	Complete the current task after verification
/taskwing:context	Get full project knowledge dump for complete architectural context

CLI verbs the slash commands rely on

The four slash commands above are pure prompts that drive these CLI verbs. You can run any of them by hand to drive TaskWing without an AI tool:

Verb	Purpose
taskwing ask "<query>" --json	Search project knowledge (decisions, patterns, constraints)
taskwing knowledge --json	Dump every knowledge node, grouped by type
taskwing task next --json	Get the next pending task
taskwing task current --json	Show the current in-progress task
taskwing task start <id>	Claim a task
taskwing task complete <id> --summary "..." --files a,b,c	Mark complete
taskwing plan --params '<json>'	Drive a plan flow (clarify → decompose → expand → finalize)

Autonomous task execution (hooks)

TaskWing integrates with Claude Code's hook system for autonomous plan execution:

taskwing hook session-init      # Initialize session tracking
taskwing hook continue-check    # Check if should continue to next task
taskwing hook session-end       # Cleanup session
taskwing hook status            # View current session state

Circuit breakers prevent runaway execution:

• --max-tasks=5 -- Stop after N tasks for human review
• --max-minutes=30 -- Stop after N minutes

AWS Bedrock setup

llm:
  provider: bedrock
  model: anthropic.claude-sonnet-4-5-20250929-v1:0
  bedrock:
    region: us-east-1
  apiKeys:
    bedrock: ${BEDROCK_API_KEY}

Model	Use case
anthropic.claude-opus-4-6-v1	Highest quality reasoning
anthropic.claude-sonnet-4-5-20250929-v1:0	Best default balance
amazon.nova-premier-v1:0	AWS flagship Nova
amazon.nova-pro-v1:0	Strong balance
meta.llama4-maverick-17b-instruct-v1:0	Open-weight general model

Or configure interactively: taskwing config

• taskwing init
• taskwing learn
• taskwing ask "<query>"
• taskwing knowledge
• taskwing task <next|current|start|complete>
• taskwing plan --params '<json>'
• taskwing doctor
• taskwing config
• taskwing start

Documentation

• Getting Started
• Product Vision
• Architecture
• Workflow Pack

License

MIT