Tachikoma

A general purpose AI agent named after the curious AI tanks from Ghost in the Shell.

Features

• Cost-Aware Routing: Match task complexity to optimal execution strategy
• PAUL Methodology: Structured Plan-Apply-Unify loop with mandatory closure
• Verification Loops: Generator-Verifier-Reviser pattern for complex tasks
• Position-Aware Context: Mitigates U-shaped attention bias in LLMs
• Model-Aware Editing: Dynamic edit format selection based on model
• MCP Integration: Leverages external Model Context Protocol servers for enhanced capabilities

Installation

bun run install.ts

Interactive installer with three options:

• Local - .opencode/ (current project only)
• Global - ~/.config/opencode/ (all projects)
• Custom - Specify any installation path

After installation, run opencode and use @tachikoma in the TUI.

See Installation Guide for detailed installation options.

Themes

Ghost in the Shell inspired themes for OpenCode terminal:

Theme	View	Dark	Light
ghost-in-the-shell	Start
lucent-ghost-in-the-shell	Start

Other screenshots here

Using Tachikoma Tools

Tachikoma exposes scripts as OpenCode tools and integrates with MCP servers. See MCP Integration for details on external tools.

@tachikoma Check edit format for current model
# → Uses tachikoma.edit-format-selector (MCP or local)

@tachikoma Query graph memory
# → Uses tachikoma-mcp_query_graph_memory (MCP with local fallback)

@tachikoma Process large context
# → Uses tachikoma-mcp_enhanced_rlm_process (MCP with local fallback)

Core Concepts

Cost-Aware Routing

Complexity	Strategy	Latency
Low	Direct response	1-2s
Medium	Single skill	5-15s
High	Skill chain	15-45s
Very High	RLM orchestration	45-120s

PAUL Methodology

1. PLAN: Define objective, acceptance criteria (Given/When/Then), tasks with verify steps, boundaries
2. APPLY: Execute tasks sequentially, each with verification
3. UNIFY: Reconcile plan vs actual, update .tachikoma/state/STATE.md, create .tachikoma/state/summary.md

Never skip UNIFY - this is the heartbeat that prevents drift.

Verification Loops

For complex implementations, use up to 3 verification iterations:

1. GENERATE - Produce initial solution
2. VERIFY - Check with explicit criteria
3. REVISE - Fix based on verification
4. REFLECT - Question approach, flag issues

Use verification for: complex implementations, high-stakes fixes, first-time features, correctness-critical tasks.

Adding New Scripts

Note: During installation, src/ is copied to .opencode/ (the OpenCode convention). Add scripts to the installed location, not src/.

1. Create a new .ts file in .opencode/plugin/tachikoma/:

#!/usr/bin/env bun
/**
 * My new capability
 * Description of what it does
 */

const args = Bun.argv.slice(2);

// Your logic here
console.log(`Processing: ${args[0] || "no args"}`);

2. Reinstall: bun run install.ts

3. Script automatically becomes tachikoma.my-new-capability tool!

Documentation

Quick Links

• Getting Started - Quick start guide
• Installation Guide - Installation options and setup
• src/agents/tachikoma.md - Agent configuration

Full Documentation

• VitePress Site
• docs/ - Raw markdown source

License

MIT