The secure AI coding
orchestrator
Point it at any project. Talk to it. It reads, edits, and commits inside a locked-down sandbox. Run locally with open models. No accounts. No costs. Your machine, your rules.
- read_file
- write_file
- apply_code_patch
- search_project
- list_project_files
- run_bash_command
- git_commit
- git_status
Your project is mounted read-write inside a secure container. No code leaves your machine except to the LLM API, or stays fully local with open models.
Tiered reasoning across Claude, Grok, GPT-4, Gemini, or run fully air-gapped with local models via vLLM on your own GPU.
Exposes a standard /v1/chat/completions endpoint. Plug in Aider, Crush, OpenCode, or any OpenAI-compatible tool.
Human-in-the-loop security for all file edits. Nothing gets written without your explicit approval.
Secure by Default
Zero-trust execution model. Sandboxed containers, approval gates, and auditable edit pipelines. The opposite of run anything and hope for the best.
Fully Local Operation
Run with open models on your own GPU via vLLM. No accounts, no API costs, no data leaving your machine. True air-gapped AI coding.
Headless Kernel
Not a UI. An intelligent backend protocol. Connect your favorite terminal UI and let Tendril handle the thinking, editing, and committing.
Self-Building
In Root Agent mode, Tendril operates on its own source code. The system that builds and repairs itself through a staged edit pipeline.
Multi-LLM Intelligence
Tiered reasoning across providers. Use the best model for each task, or fall back gracefully. Claude, Grok, GPT-4, Gemini, or fully local.
MIT Licensed
No restrictive licensing. Build freely, fork freely, deploy freely. The moat is velocity and the managed experience, not the code.
Aider
The gold standard for Git-native code editing.
export OPENAI_API_BASE="http://localhost:8080/v1" export OPENAI_API_KEY="tendril" aider --model openai/tendril
Crush
Glamorous, blazing-fast Terminal UI experience.
export OPENAI_API_BASE="http://localhost:8080/v1" export OPENAI_API_KEY="tendril" crush --model openai/tendril
OpenCode
Neovim-centric, highly customizable agenting.
# Point OpenCode's custom provider # configuration to: http://localhost:8080/v1
# Clone and configure git clone https://github.com/opentendril/core.git cd core && cp .env.example .env # Add your LLM API key (or configure local model) # Then launch: docker compose up --build # Verify it is live: curl -s http://localhost:8080/health | python3 -m json.tool # - {"status": "healthy", ...}
Evolution 1: The Root Agent
Self-healing, self-building. Multi-LLM routing, approval gate, the /edit endpoint loop, and the Chronicler. Tendril pointed at its own source code.
Evolution 2: Deployed Agents
Pointing Tendril at external codebases. External project mode, dynamic workspace awareness, tool adaptability. Enterprise pipelines and local-first compute coming next.
Evolution 3: The Automated Hive
Sharing skills learned by deployed agents back to the Root. Skill abstraction, automated PRs to a global registry, federated learning across Tendril instances.
Livestream Coming Soon
The self-building kernel, live and unfiltered. Star the repo to get notified.
Star on GitHubGet Involved
Or send a direct message.