CUDA production
asm-agentAuthoritative local ASM-CM inference. Requires CUDA-ready PyTorch and an NVIDIA GPU.
Terminal agent · RLM-native
ASM Agent
Build with an agent that can inspect its own work.
ASM Agent is an RLM-native terminal harness for coding and research. It gives the model a persistent Python kernel, explicit tools and a workflow built around execution, observation and iteration.
Install ASM Agent ↓System requirements
Know what runs locally before you install.
ASM Agent is local software. The verified installer downloads the release, prepares isolated Python environments and installs the bundled ASM-CM checkpoint. Internet access is required during initial setup; a checkpoint install typically takes 5–20 minutes.
| Component | Minimum | Production default |
|---|---|---|
| Operating system | Linux, macOS, or Windows through WSL | Linux or WSL with a supported NVIDIA driver |
| Node and npm | Node.js 22.8+ and npm 11.10+ | Node.js 22 LTS with npm 11.10 or newer |
| Python | Python 3.12–3.14 and uv | Installer-managed Python 3.12 environments |
| Database | PostgreSQL 17 or 18 | PostgreSQL 17/18; Docker Compose is optional |
| GPU | No GPU with explicit CPU diagnostics or ASM-CM disabled | CUDA-capable NVIDIA GPU; startup fails closed without CUDA |
| Storage | About 1 GiB for the CPU runtime and checkpoint | Minimum plus the selected CUDA/PyTorch wheel and runtime caches |
| RAM / VRAM | No fixed validated minimum is published | Size for the selected PyTorch/CUDA runtime and verify with doctor |
| Network | Required for initial installation | Required for setup, updates, and remote model providers |
Execution profiles
After installation, verify the local runtime with asm-agent doctor --json.
CPU
ASM_AGENT_ASM_CM_DEVICE=cpu asm-agentExplicit CPU diagnostics. It is not the default production profile or the automated full-training path.
ASM-CM disabled
ASM_AGENT_ASM_CM_MODE=disabled asm-agentStarts without ASM-CM for an explicit fallback session. Neural-state features are unavailable.
01 · Install ASM Agent
Install ASM Agent
Run the installer on macOS, Linux or Windows through WSL. The first launch prepares the Python kernel automatically.
curl -fsSL https://aletheionagi.com/asm-agent/install.sh | shChatGPT or Claude Code
Authentication is available only through ChatGPT and Claude Code login.
Pretrained assets
The standard installer downloads pretrained model assets. No local training is required.
Then start it in your project:
$ asm-agent02 · Core capabilities
One terminal.
A controlled execution loop.
Persistent kernel
Python state survives between steps, so the agent can keep findings, handles and intermediate results while it works.
RLM-native workflow
The model breaks work into steps, executes code, observes results and changes course when evidence requires it.
Extensible tools
Add skills, MCP integrations, prompt templates, extensions and reusable ASM Agent packages.
Supported login
Authentication is available only through ChatGPT and Claude Code login.
03 · Designed for real repositories
Inspect. Execute. Verify.
ASM Agent works inside your project instead of guessing from an isolated chat. It can read files, run project commands, edit code and inspect data through a persistent environment. You remain in control of the repository, credentials and changes.