📚 Related: Best Models for Coding Locally · Qwen Models Family Guide · Ollama vs LM Studio · The 5 Levels of AI Coding · VRAM Requirements

Claude Code with Opus is the best AI coding tool available. It’s also $100/month for Claude Max 5x or $200/month for Max 20x (Anthropic’s current rate card as of June 2026), sends your code to Anthropic’s servers, and requires an internet connection.

If you have a 24GB GPU and run Ollama, you can get surprisingly close with open-source tools and local models. Not all the way — frontier models still handle complex multi-file refactors better than anything running on consumer hardware. But for tab completion, single-file edits, bug fixes, and routine coding tasks, local is genuinely practical in mid-2026.

Here’s what works, what doesn’t, and how to set up the best local coding stack.

What’s New (June 2026)

“Claude code alternative open source” is the #1 query our demand analyzer is tracking — 5,738 standing demand, climbing week-over-week. The article below still maps the original 2026 contenders (Aider, Continue.dev, Cline, OpenCode), and that survey is still useful. Here’s what’s changed since February.

OpenClaw has emerged as a major Claude Code alternative. It’s not in the table below because it didn’t exist in this category three months ago. Active community, frequent releases, and a broad skill ecosystem. Start with the OpenClaw Setup Guide and the Best Models for OpenClaw writeup — between them they’ll get you to a working agent in an evening.

Qwen 3.6 changed the model substrate underneath all of these tools. Qwen 3.6-27B dense at Q4_K_M (~17 GB) is the new recommended local backend for OpenClaw and the article’s existing contenders. SWE-bench Verified at 77.2 — close enough to frontier for serious agentic coding, on a single 24GB card. The body recommendations and setup commands below have been updated accordingly, replacing the earlier Qwen 2.5 Coder 32B framing. Full breakdown: Qwen 3.6 Complete Guide.

DFlash + DDTree decode speedup is real and reproducible. Speculative decoding for the 27B dense — mainline MTP support landed in llama.cpp (PR #22673), and the Luce-Org DFlash fork lifts an RTX 3090 to ~2.56x mean decode throughput on Qwen 3.6-27B (firsthand bench, April 30). For agentic workloads where you’re waiting on token generation, that’s the difference between “usable” and “fast.” Comparison: DFlash vs MTP head-to-head.

PI Agent is now covered in a dedicated section below — Mario Zechner’s MIT-licensed terminal harness makes a minimal counterpoint to the more feature-heavy options.

The open-weight backend field expanded. Three notable model releases since the May refresh — none of them runs-on-consumer-hardware picks at full size, all relevant to the local-vs-API decision for the harnesses below:

• GLM-5.2 (Z.ai, released June 13, 2026): MIT-licensed 744B MoE with 40B active per token, 1M context, tops the open-weights Intelligence Index at 51 (coverage). Integrates with Claude Code, Cline, OpenCode, and Roo Code via the GLM Coding Plan. API or datacenter-scale, not a 24GB-card model — listed here for backend awareness, not as a local pick.
• DeepSeek V4-Flash and V4-Pro: the price play. V4-Flash at $0.14/$0.28 per million tokens (cache-miss input/output), V4-Pro at $0.435/$0.87 after the May 31 permanent price cut (75% off launch promo made permanent — pricing detail). V4-Flash is the agentic-tool-call pick, V4-Pro the heavy-reasoning option. Full breakdown: DeepSeek V4 Flash vs Pro guide.
• Kimi K2.6 and K2.7-Code (Moonshot AI): K2.6 (April 2026) ships open-weights agentic coding with 300-agent swarms and 4,000-step horizons; K2.7-Code (June 12, 2026) is the newer coding-specialized variant with ~30% fewer thinking tokens. Open weights on Hugging Face, datacenter or rented-GPU territory at full size.

For the full local-coding model field by VRAM tier including Gemma 4 26B-A4B MoE, the Qwen 3.6 lineup in depth, and the model-pick decisions, see Best Local Coding Models 2026 — that’s the sibling page that owns model picks. This page focuses on the harness/tool decisions.

The framework below — local for the 80% routine, frontier for the hard 20% — still holds. Just substitute Qwen 3.6 for Qwen 2.5 Coder, and add OpenClaw to the agent shortlist.

The Local Backend Underneath These Tools

The harness sections below assume Qwen 3.6 (ollama pull qwen3.6:35b for the 35B-A3B MoE on 24GB cards, or ollama pull qwen3.6:27b for the dense 27B coder) as the starting backend. For the full local-coding model field by VRAM tier — Qwen 3.6 in depth, Gemma 4 26B-A4B MoE, Qwen 2.5 Coder 7B for FIM tab completion, the 80B-A3B picks for unified-memory setups — see Best Local Coding Models 2026. That sibling page owns model picks; this page focuses on which harness/tool to point at whichever backend you chose.

ollama pull qwen3.6:35b

Aider — Best Terminal Agent

Stars: 45,400 | License: Apache 2.0 | GitHub

Aider is the closest thing to Claude Code that runs with local models. You run it in a git repo, describe changes in natural language, and it edits your files directly with automatic git commits.

Setup

pip install aider-chat
ollama pull qwen3.6:35b
aider --model ollama/qwen3.6:35b

Strengths

• Builds a repo map of your entire codebase for context
• Automatic git integration — every change is a commit you can undo
• Voice coding support
• Mature (since 2023) with transparent model benchmarks
• Works with 90+ languages

Weaknesses

• Terminal-only (no GUI; third-party Aider Desk wrapper exists)
• Local models struggle with large multi-file refactors compared to Claude/GPT
• No browser automation or MCP support
• Repo map generation is slow on very large codebases

Verdict

The best terminal-based coding agent for local models. If you liked Claude Code’s terminal workflow but want to run on your own GPU, start here.

Continue.dev — Best IDE Tab Completion

Stars: 33,400 | License: Apache 2.0 | GitHub

An open-source VS Code / JetBrains extension that provides tab completion and inline chat. Think “open-source Copilot” wired to any backend. As of mid-2026 the project has expanded into CI-enforceable AI checks via a Continue CLI, but the VS Code extension remains the most common entry point for local-model users.

Setup

1. Install “Continue” from VS Code marketplace
2. In ~/.continue/config.yaml (the config.json format is deprecated), add Ollama as a provider:
   • Chat model: qwen3.6:35b
   • Tab completion model: qwen2.5-coder:7b (smaller, faster, code-specialized for FIM)

Strengths

• Lives inside your editor — no context switching
• Tab completion feels like Copilot with a fast local model
• Codebase indexing for RAG-style context retrieval
• Can mix local and cloud models (local for autocomplete, Claude for complex tasks)

Weaknesses

• Not an agent — it doesn’t execute commands, create files, or run tests autonomously
• Tab completion quality with 7B models is noticeably below Copilot for complex completions
• Configuration can be fiddly (model parameters, prompt templates, context windows)

Verdict

The best way to get Copilot-style tab completion running entirely on your GPU. Pair it with Aider or Cline for agent-style tasks.

Cline — Best VS Code Agent

Stars: 62,400 | Installs: 5M+ | License: Apache 2.0 | GitHub

The most-installed open-source AI coding agent for VS Code. Originally “Claude Dev.” Supports Plan/Act modes, MCP integration, file editing, terminal commands, and browser automation — all with explicit user approval at each step.

Setup

1. Install “Cline” from VS Code marketplace
2. Set provider to Ollama, select your model
3. Every action requires your approval (approve/deny per tool call)

Strengths

• Full agent capabilities (file create/edit, terminal, browser)
• Explicit approval workflow prevents surprises
• MCP integration for custom tools
• Cline CLI 2.0 brings it to the terminal with parallel agents

Weaknesses

• Agent loop with local models fails more often than with Claude/GPT
• Approval-per-action gets tedious on long tasks
• Heavy token consumption

Verdict

Best VS Code agent if you want autonomous capabilities with safety rails. Works with local models, but expect more iterations than with frontier models.

OpenCode — Go-Based Terminal Agent

Stars: ~12,700 | License: MIT | GitHub

An open-source terminal coding agent written in Go with a TUI, multi-session support, LSP integration, and compatibility with 75+ models including local ones. Last active development was September 2025, so treat the project as functional but quiet.

Setup

# Install via Go
go install github.com/opencode-ai/opencode@latest
# Or download binary from GitHub releases
opencode --provider ollama --model qwen3.6:35b

Strengths

• Fast (Go binary, not Python)
• Multi-session support
• IDE extensions for VS Code, JetBrains, Neovim, Zed, and Emacs
• GitHub Actions integration via /opencode comments
• Agent Client Protocol (ACP) for editor communication

Weaknesses

• Primarily designed around cloud models — local model support works but isn’t the primary focus
• Repository has been quiet since September 2025; bug fixes and model-compatibility updates are not landing on the upstream

Verdict

Worth trying if you want a fast, Go-based terminal agent and you’re comfortable with a stable-but-quiet upstream. The multi-editor support via ACP is a genuine differentiator.

PI Agent — Minimal Terminal Harness

Stars: 55,500 | License: MIT | GitHub

Mario Zechner’s MIT-licensed terminal coding agent — built around aggressive minimalism: ~200-token system prompt, four default tools (read, write, edit, bash), and YOLO-by-default execution. Extended through TypeScript skills, prompt templates, and pi packages rather than fiddly configuration.

Setup

npm install -g --ignore-scripts @earendil-works/pi-coding-agent
# Or use the official installer:
curl -fsSL https://pi.dev/install.sh | sh
ollama pull qwen3.6:35b
pi

Configure providers in ~/.pi/agent/models.json. The npm package moved from @mariozechner/pi-coding-agent to the current path in May 2026; same authors, same binary, same config schema.

Best fit

Qwen 3.6 35B-A3B on a 24GB GPU (or 16GB with --cpu-moe offload). The minimal harness rewards a capable model — full setup walkthrough in Best Local Models for PI Agent.

Void — Open-Source Cursor

Stars: 28,800 | License: MIT | GitHub | Y Combinator backed

An open-source VS Code fork that aims to replicate Cursor’s feature set. Agent mode, inline editing, contextual chat.

Setup

Download from voideditor.com. It auto-detects Ollama at http://127.0.0.1:11434 and transfers your VS Code themes, keybinds, and settings in one click.

Strengths

• Closest open-source equivalent to Cursor
• Full VS Code extension compatibility
• No middleman server — connects directly to Ollama
• You can view and edit the prompts sent to the AI

Weaknesses

• Still in beta — expect bugs and rough edges
• Agent mode with local models is significantly less capable than with Claude/GPT
• Smaller team than Cursor

Verdict

Best option if you want a Cursor-like experience with local models. Still maturing.

Tabby — Best for Teams

Stars: 33,500 | License: Apache 2.0 | GitHub

Self-hosted coding assistant server. Run a Tabby server on your hardware, get code completion and chat in VS Code, JetBrains, or Vim.

Setup

docker run -it --gpus all -p 8080:8080 \
 tabbyml/tabby serve --model StarCoder-1B --chat-model Qwen2-1.5B-Instruct

Strengths

• Purpose-built for self-hosting and enterprise deployment
• Repository-level code indexing (connects to GitHub, GitLab, local repos)
• Multi-user support with admin dashboard
• Clean REST API

Weaknesses

• Primarily completion and chat, not an autonomous agent
• No terminal/CLI agent mode
• Smaller model ecosystem than Continue

Verdict

Best choice for teams that want a self-hosted Copilot with admin controls and repository indexing.

Quick Comparison

Where Local Closes the Gap (and Where It Doesn’t)

Local models are competitive for:

• Tab completion: Qwen 2.5 Coder 7B running locally feels snappier than cloud Copilot due to zero latency
• Single-file edits: 32B models handle “add a function” and “fix this bug” competently, trailing current frontier API models (Claude Opus 4.8, GPT-5.2) only on harder edits
• Privacy: The only option for air-gapped environments and proprietary codebases
• Cost at scale: Free after GPU investment vs $100-$200/month for Claude Max

The gap remains significant for:

• Multi-file refactoring: Claude Opus with 200K context can coordinate changes across dozens of files. Local 32B models degrade past 32K tokens in practice
• Complex architectural reasoning: Frontier models suggest patterns and trade-offs that 32B models cannot
• Agent loop reliability: Cloud models complete agent tasks in fewer iterations with fewer failures
• Token efficiency: Claude Code uses 5.5x fewer tokens than Cursor for identical tasks. Local models are even less efficient

The Recommended Setup

Hardware tier — Model

 24GB GPU: Qwen 3.6 35B-A3B (clean MoE on a single RTX 3090 / 4090)
 16GB GPU: Qwen 3.6 35B-A3B with --cpu-moe (routed experts on system RAM)
 8GB GPU: Qwen 2.5 Coder 7B for tab completion / FIM
 Qwen 3.5 9B for chat and small edits

Tools:
 - Continue.dev (or LM Studio with MLX on Mac) for tab completion
 - Aider, Cline, PI Agent, or OpenClaw for agent-style editing
 - Frontier model via API for hardest cases — Claude Opus 4.8, GPT-5.2,
 or DeepSeek V4-Flash through the DeepSeek API

→ Not sure what fits? Try our Planning Tool.

This hybrid approach — local for the 80% of routine work, cloud for the 20% of hard problems — is the most cost-effective setup in mid-2026. The local tools are genuinely good enough for daily coding. They’re just not quite good enough to replace frontier models on the tasks where you most want help.

That gap is narrowing. Check back in six months.

Get notified when we publish new guides.