Open Source • Apache 2.0

Your Code.
Your Models.
Your Control.

Shofer is a locally-run, state-of-the-art, fully configurable AI coding agent that operates as a VS Code extension — running entirely on your machine, under your control.

Star on GitHub Install from VS Code Marketplace

Runs locally Parallel tasks Any model 50+ tools

Features

Everything you need to code at the speed of thought

Shofer packs the most comprehensive tool set of any open-source AI coding agent — from semantic code search to parallel task orchestration.

5 Built-in Modes

Switch between Code, Architect, Ask, Debug, and Orchestrator modes — each with scoped tool access. Create fully custom modes via .shofermodes files to match your exact workflow.

Fine-grained tool category control (read, write, execute, mcp, etc.)
File-scoped restrictions (e.g., write only .md files)
Custom role definitions and model assignments per mode

Custom Modes

True Parallel Tasks

Run multiple independent conversations simultaneously in a tree hierarchy. Each task has its own mode, provider, and context — switch freely without losing state.

Background subtasks with parent-child communication
Message queuing — type ahead while the AI works
Task states with colored indicators (idle, running, waiting, completed)

Parallel Tasks

Git Worktrees

Run parallel tasks on different branches in one VS Code window. No more stash/commit gymnastics or multiple windows for PRs.

UI-driven create, switch, and delete operations
Each task scoped to a specific worktree branch
Automatic .worktreeinclude for gitignored files

Git Worktrees

RAG Code Indexing

Build a semantic search index of your codebase and git history. Find code and commits by meaning — not just keywords.

AST-aware parsing via tree-sitter for accurate code chunking
Git history semantic search — discover why and when changes happened
File watcher keeps the index up to date as you edit

RAG Indexing

Assistant Agent

A persistent, read-only AI companion that accumulates codebase knowledge across tasks — surviving restarts. Answers questions without reloading context.

Runs on a low-cost model with a large context window
KV-cache friendly append-only context minimizes provider costs
Strictly read-only — can read, search, and look up symbols only

Assistant Agent

Async MCP Tool Calling

Fire-and-forget MCP tool invocation with true parallelism. Fan out multiple calls and collect results when they're ready.

Non-blocking call_mcp_tool_async with call_id tracking
wait_for_mcp_call with all/any completion strategies
Delete-on-read trimming prevents context bloat

MCP Servers

Modes

5 built-in modes. Infinite customizations.

Each mode has scoped tool access. Create your own via .shofermodes files — project-level or global.

💻

Code

Writing, modifying, and refactoring code. Broadest tool access.

read, write, execute, mcp, mode, subtasks, questions

🏗️

Architect

Planning and designing before writing code. Read + markdown-only writes.

read, write (.md only), mcp, questions

❓

Ask

Getting explanations, answers, or recommendations. Read-only + MCP.

read, mcp

🪲

Debug

Troubleshooting errors and diagnosing root causes.

read, write, execute, mcp, subtasks, questions

🪃

Orchestrator

Coordinating complex multi-step work by delegating to sub-tasks.

delegates via new_task

Common custom modes include Reviewer, Search, Opinion, and Browser.

Learn how to create custom modes

Migration

Already using another AI coding tool?

One-command migration from Roo-Code and GitHub Copilot. Your configuration, modes, and rules come with you.

Migrating from Roo-Code

Shofer is a major architectural improvement over Roo-Code — a complete rebuild with parallel tasks, async MCP calling, semantic code & git indexing, native worktree support, and dozens of new features.

/migrate-from-roocode Full guide ↗

Feature	Shofer	Roo-Code
Parallel Tasks	Multiple concurrent independent conversations with task tree hierarchy	Single task at a time — starting a new task abandoned the current one
Background Subtasks	Fan out work without blocking the parent. Parent can answer child questions, cancel children, and poll results	Synchronous only — parent blocked until child completed
Async MCP Calls	Fire-and-forget with call_id tracking, all/any wait strategies, delete-on-read trimming	All MCP calls were synchronous and blocking
RAG Code Indexing	Fast AST-aware tree-sitter parsing, git history semantic search, submodule-aware scanning	Limited to code, and slow
Git Worktrees	Native UI-driven worktree management within one VS Code window	Required separate VS Code windows per worktree
Cost Limits	Set a USD budget cap per task globally and per-task. When reached, the task automatically pauses or aborts — preventing runaway spend.	No spending limits — runaway loops could consume tokens indefinitely
Assistant Agent	Persistent read-only companion with cross-session context reuse	Not available
LSP Symbol Refactoring	rename_symbol and list_code_usages via Language Server Protocol	No LSP-powered refactoring tools

Migrating from GitHub Copilot

Shofer offers everything Copilot does and much more — with full privacy, model autonomy, and granular control over every aspect of your AI coding experience.

/migrate-from-copilot Full guide ↗

Aspect	Shofer	Copilot
Data Processing	100% local execution — no remote server processes your workspace	Cloud-proxied through Microsoft/GitHub infrastructure
Model Ecosystem	Bring-Your-Own-Model — Ollama, LM Studio, Anthropic, OpenAI, OpenRouter, and more	Curated cloud-managed multi-model picker
Network Autonomy	Fully operational in air-gapped environments with local models	Requires active internet and constant telemetry handshake
Agent Orchestration	Agent-to-Agent orchestration — spawn, monitor, and converse with child agents	Monolithic single-agent — no dynamic sub-agent management
Async MCP	Fire-and-forget async tool calling with parallel execution	Synchronous request-response — entire loop blocks on tool calls
Cost Control	Per-task USD cost caps with automatic agent halting	No per-session cost tracking or limits
Subscription	Free Open Source (Apache 2.0) — pay only for your own API tokens	$10–$39/month SaaS subscription