Installation

AdaMAST supports a dependency-light base install plus optional model-provider extras.

Requirements

• Python 3.10 or newer
• A project directory where AdaMAST can write local hook config, if you use a harness integration
• A writable trace directory

Codex- and Claude-native taxonomy learning use the active conversation's subagent capability and need no standalone host CLI or second login. Explicit provider-backed project configurations still require the selected provider's normal credentials.

Install from PyPI

python -m pip install adamast

Install from GitHub (source)

python -m pip install "git+https://github.com/multi-agent-systems-failure-taxonomy/ATLAS.git"

Install from a local checkout

cd /path/to/AdaMAST
python -m pip install .

For editable development:

python -m pip install -e ".[test]"

Zero-config interactive install

Install AdaMAST once for every project in a host:

adamast-codex-install --user-level
adamast-doctor --codex

adamast-claude-install --user-level
adamast-doctor --claude-code

These defaults use ~/.adamast/interactive, automatic Git-project scoping, and the browser selector. Codex and Claude Code use a native subagent in the active task for taxonomy learning and do not require a standalone host CLI login or separate provider API key. Run both installers to share project/task-group state across both hosts.

See Interactive setup for behavior and uninstall steps.

Optional provider extras

Anthropic SDK:

python -m pip install "adamast[anthropic]"

AWS Bedrock Converse through boto3:

python -m pip install "adamast[bedrock]"

For Bedrock, set credentials in the environment:

export AWS_BEARER_TOKEN_BEDROCK="..."
export AWS_REGION="us-east-1"

AdaMAST reads provider credentials from the environment. Do not put tokens in adamast.json.

Minimal provider-backed project config

Create adamast.json in the project using AdaMAST:

{
  "version": 1,
  "trace_output": "./adamast-program",
  "trace_root": "~/.adamast/traces",
  "store_dir": "~/.adamast/taxonomies",
  "adamast_model": "gpt-5",
  "inherit": null,
  "generation_threshold": 5,
  "generation_stops": false,
  "skip_judge": false,
  "k_init": 10,
  "k": 20,
  "refinement_stops": false,
  "advanced_refinement": false,
  "freeze": false,
  "repair_rounds": 3,
  "format_retries": 2,
  "gate_exhaustion_policy": "raise",
  "recent_activity_messages": 8,
  "recent_activity_chars": 12000,
  "dashboard": true,
  "claude_code": {
    "built_in_hooks": {
      "SubagentStop": true,
      "PostToolUse": ["Bash", "Edit", "Write"]
    },
    "custom_hooks": []
  },
  "codex": {
    "hooks": {
      "SessionStart": true,
      "Stop": true,
      "SubagentStop": true,
      "PostToolUse": ["shell_command", "apply_patch"]
    }
  }
}

Relative paths are resolved relative to the config file.

Use adamast_model for AdaMAST generation, checking, and refinement calls. If your task-solving program also has a model flag, keep it separate.

Verify the install

adamast-doctor --config adamast.json

For harness-specific checks:

adamast-doctor --config adamast.json --claude-code
adamast-doctor --config adamast.json --codex

Errors mean the requested setup is not ready. Warnings usually mean AdaMAST can run, but an optional integration or dependency is incomplete.