TaskWing

Your AI assistant forgets everything. Every single session.

Context compression tools save tokens. TaskWing saves knowledge — decisions, patterns, and architecture that compound across every session.

Website · Tutorial · Vision · Install

Works With

---

The Problem

You explain "we use PostgreSQL, not MongoDB" on Monday. Again on Tuesday. Again on Wednesday. Your AI assistant has no memory. Every session starts from zero.

A typical project accumulates 50+ architectural decisions, dozens of patterns, and critical constraints — none of which survive a session restart. You spend more time re-explaining context than building features.

Context compression tools reduce token waste within a session. But when the session ends, everything is gone. The real problem isn't token cost — it's knowledge loss.

How TaskWing Fixes This

Without TaskWing:
  Session 1: "We use PostgreSQL, here's why..." (re-explain)
  Session 2: "We use PostgreSQL, here's why..." (re-explain)
  Session 3: "We use PostgreSQL, here's why..." (re-explain)

With TaskWing:
  taskwing bootstrap → 63 decisions, 28 patterns, 12 constraints extracted
  Session 1: AI already knows your architecture
  Session 2: AI still knows. Plus what you decided yesterday.
  Session 3: AI knows everything. Context compounds.

One command extracts your architecture into a local SQLite database. Every AI session after that just knows — permanently.

What	How
Persistent memory	Decisions, patterns, and constraints survive every session restart
Goal-to-execution	Turn "Add Stripe billing" into 5 decomposed, context-rich tasks
Code intelligence	Symbol search, call graphs, impact analysis across your codebase
Works everywhere	Claude Code, Cursor, Copilot, Gemini, OpenCode — via MCP

Try It (60 seconds)

# Install
brew install josephgoksu/tap/taskwing

# Extract your architecture
cd your-project && taskwing bootstrap
# → "63 decisions, 28 patterns, 12 constraints extracted"

# Ask it anything about your project
taskwing ask "what database do we use and why?"
# → Returns the decision, reasoning, and tradeoffs — instantly

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

Full Workflow

# 1. Bootstrap (once per project)
taskwing bootstrap

# 2. Set a goal and generate a plan
taskwing plan "Add Stripe billing"
# → Plan decomposed into 5 executable tasks with architectural context

# 3. Execute with your AI assistant
/tw-next       # Get next task — AI already knows your stack
# ...work...
/tw-done       # Mark complete, advance to next

# 4. Your AI remembers decisions made today — tomorrow, next week, forever
/tw-ask "why did we choose Stripe over Paddle?"
# → Returns the decision from step 2, with full reasoning

Supported Models

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

Why Not Just a CLAUDE.md File?

	CLAUDE.md	Context compression tools	TaskWing
Survives session restart	Yes	No	Yes
Auto-extracted from code	No (hand-written)	No	Yes
Searchable	No	Session only	Always (FTS + vector + graph)
Grows over time	Only if you maintain it	No	Automatically
Understands code symbols	No	No	Call graphs, impact analysis
Plans and tracks tasks	No	No	Goal → plan → execute → verify

CLAUDE.md is a good start. Context compression is useful within a session. TaskWing is what happens when your project intelligence becomes permanent, searchable, and compounding.

MCP Tools

Tool	Description
ask	Search project knowledge (decisions, patterns, constraints)
task	Unified task lifecycle (next, current, start, complete)
plan	Plan management (clarify, decompose, expand, generate, finalize, audit)
code	Code intelligence (find, search, explain, callers, impact, simplify)
debug	Diagnose issues systematically with AI-powered analysis
remember	Store knowledge in project memory

MCP Setup

Add to your AI tool's MCP config:

{
  "mcpServers": {
    "taskwing-mcp": {
      "command": "taskwing",
      "args": ["mcp"]
    }
  }
}

Slash Commands

Once connected, use these slash commands from your AI assistant:

Command	When to use
/tw-ask	Search project knowledge (decisions, patterns, constraints)
/tw-remember	Persist a decision, pattern, or insight to project memory
/tw-next	Start the next approved task with full context
/tw-done	Complete the current task after verification
/tw-status	Check current task progress and acceptance criteria
/tw-plan	Clarify a goal and build an approved execution plan
/tw-debug	Root-cause-first debugging before proposing fixes
/tw-explain	Deep explanation of a code symbol and its call graph
/tw-simplify	Simplify code while preserving behavior

Core Commands

• taskwing bootstrap
• taskwing plan "<description>"
• taskwing ask "<query>"
• taskwing task
• taskwing plan status
• taskwing slash
• taskwing mcp
• taskwing doctor
• taskwing config
• taskwing start

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

TaskWing supports Bedrock as a first-class provider:

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

Recommended Bedrock models

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

Documentation

• Getting Started
• Product Vision
• Architecture
• Workflow Contract v1
• Workflow Pack

License

MIT