48 tools  ·  Cloud Models  ·  One command install

Your AI coding agent, cloud-powered & private.

Terminal-first AI coding agent powered by DeepSeek. 48 tools. Understands your codebase. Fixes bugs. Ships features. Uses Threadlane's cloud models — no GPU required.

macOS & Linux
No GPU needed
Free & open source
threadlion-code — ~/my-project
All 48 built-in tools
48Built-in toolsread, write, search, memory
0API keys neededmodels on Threadlane cloud
<5sFirst responsecloud inference, instant start
1Command installcurl | bash and you're done
The problem

Great AI tools exist.
They just send your code to the cloud.

Every keystroke in Copilot, every paste in ChatGPT, every file in Cursor — leaves your machine. ThreadLion Code keeps it all local.

Without ThreadLion Code
  • Code sent to OpenAI, Anthropic, or MS servers
  • API bills that scale with usage
  • No audit trail of what was sent
  • Rate limits break your workflow
  • Context window resets every chat
With ThreadLion Code
  • All inference on your hardware via Ollama
  • Flat cost — pay once for hardware
  • Full repo indexed in persistent memory
  • No rate limits — run as many sessions as you want
  • Session history persists across restarts
You stay in control
  • Shell command approval gate — nothing runs without you
  • Confidence gate — agent must be ≥70% sure before any edit
  • Syntax validation after every file write
  • Diff preview before any destructive change
  • Full harness sensor system for security + complexity
How it works

One objective. 48 tools. One answer.

The agent reads your code first, forms a plan, asserts confidence, then and only then — writes. Every time.

01
Explore first

outline_directory → summarize_file → find_symbol → find_references. The agent maps the codebase before touching a single file.

Mapping package structure... Reading auth.py — 89 lines... Found symbol: token_refresh() at line 42
02
Assert confidence

diff_preview shows what would change. assert_understanding gates on ≥70% confidence. Below that — more context, not a guess.

Diff preview — 4 lines changed assert_understanding: 91% confidence Gate passed — proceeding
03
Write and verify

edit_file or write_file, then immediately validate_source_syntax. If it fails — fix it. No marking done until syntax is clean.

Writing fix to auth.py:67... Validating syntax — clean. Done. Zero errors.
04
You approve shell commands

Every run_shell invocation waits for your explicit yes/no. Tests, linters, git — nothing executes without consent.

Run: pytest tests/auth_test.py? [y/n] y 12 passed · 0 failed
The honest comparison

Why not just use Copilot?

Capability ThreadLion Code
from $0
GitHub Copilot
$10–$39/mo
Cursor
$20–$40/mo
Runs 100% locally
No API keys needed
Code never leaves machine
Full codebase memoryPartial
Shell command execution
48 built-in toolsFewer
Multi-agent team mode
No rate limits
Monthly cost per developer$0 – $29$10 – $39$20 – $40
Pricing — your hardware, your rules

One flat price.
No per-seat trap.

Community is free and open source forever. Pro and Studio unlock team mode, admin, and priority support. All tiers run locally on Ollama — no cloud bill ever.

Community
$0 /month
Free forever · Open source
Full agent, all 48 tools, runs locally. Self-serve, community support.
  • All 48 built-in tools
  • DeepSeek + any Ollama model
  • Session + repo memory
  • Harness safety system
  • Community support (GitHub)
Most Popular
Pro
$29 /month
$23/mo billed annually — save 20%
Multi-agent team mode, admin dashboard, and priority support for solo devs who ship fast.
  • Everything in Community
  • Multi-agent team mode (Atlas + 10 workers)
  • Admin dashboard & usage analytics
  • New features 30 days early
  • Priority email support
Studio
$79 /month
$63/mo billed annually — save 20%
For dev teams. Custom agent config, white-label, SLA, and a dedicated channel.
  • Everything in Pro
  • Custom agent configuration
  • White-label (remove branding)
  • SLA guarantee
  • Dedicated Slack channel + onboarding call
Annual billing saves 20%  ·  Cancel anytime  ·  Community tier is free forever
Installation

One command. Up in 30 seconds.

The installer configures cloud models automatically — no GPU, no Ollama setup, no API keys.

Fastest: Install with Threadlane cloud models RECOMMENDED

No Ollama installation, no GPU needed. Uses code-models.threadlane.in for inference.

# Copy and paste into your terminal $ OLLAMA_HOST=https://code-models.threadlane.in curl -fsSL https://raw.githubusercontent.com/vicky4github/threadlion-code/main/install.sh | bash
Download install.sh
1
Install Ollama

Ollama runs large language models on your machine. Free and open source. Download from ollama.com or use Homebrew.

# macOS $ brew install ollama # Linux $ curl -fsSL https://ollama.com/install.sh | sh
2
Pull the AI model

Pull DeepSeek v4 Pro — the primary model. First pull downloads ~15 GB. Subsequent starts are instant.

$ ollama pull deepseek-v4-pro:cloud
3
Run the one-line installer

Downloads the binary, configures cloud models, adds threadlion to your PATH. No GPU setup needed.

$ OLLAMA_HOST=https://code-models.threadlane.in \
  curl -fsSL https://raw.githubusercontent.com/vicky4github/threadlion-code/main/install.sh | bash
4
Start the agent

Point it at your project. The agent indexes it, loads memory, and waits for your first message.

$ threadlion-code --cwd /path/to/your/project threadlion · deepseek-v4-pro · 24 files indexed You › _
install.sh
╔════════════════════════════════╗
║ ThreadLion Code Installer ║
╚════════════════════════════════╝

✓ Python 3.13.1 found
✓ Ollama: installed
✓ Ollama: started
✓ Model deepseek-v4-pro: ready
✓ ThreadLion Code: installed
✓ PATH: updated

ThreadLion Code installed!

Start the agent:
$ threadlion-code

With a project:
$ threadlion-code --cwd ~/app

Use any Ollama model:
$ threadlion-code --model llama3.3:70b
Requirements
💻 macOS or Linux
🐍 Python 3.10+ (auto-installed)
🦙 Ollama (auto-installed)
💾 ~15 GB disk for model weights
Questions

Straight answers.

No. All inference happens locally on your hardware via Ollama. The model runs on your machine, context is processed locally, and nothing is transmitted to any external server. Not to Anthropic, not to OpenAI, not to us.
For DeepSeek v4 Pro, you need at least 16 GB RAM (24 GB+ recommended for smooth performance). Apple Silicon Macs (M1/M2/M3/M4) handle this very well with unified memory. The model runs at full speed on Metal. Intel/AMD with an NVIDIA GPU also works great via CUDA.
Yes. ThreadLion Code works with any model available on Ollama — Llama 3.3, Mistral, Phi-4, Gemma 2, Qwen 2.5, and more. Pass --model model-name at startup. Quality varies by model; DeepSeek v4 Pro gives the best results for coding tasks.
The Community tier is fully open source under MIT license. You can read every line, fork it, and run it however you like. The Pro/Studio tiers add features (team mode, admin, support) — the core agent is always open.
Claude Code and Aider both require cloud API keys and send your code to external servers. ThreadLion Code runs the entire stack locally — model, memory, and tools. It also has the Understand-First protocol (explore before write), a harness safety system with confidence gates, and multi-agent team mode — all unavailable in those tools.
License verification requires a brief connection at startup. Inference, tool execution, and memory are all fully local. If you lose internet mid-session, the agent continues working normally.
Reviewed personally by the team

Your local AI coding agent
starts today.

One command. Your code stays on your machine. No API key, no cloud bill, no telemetry.

View on GitHub
Free forever on Community
Cancel Pro anytime
Annual saves 20%