How holobiomicslab/mimosa-ai works
Mimosa-AI turns a scientific objective into an executable, task-specific multi-agent workflow, runs it with discovered MCP tools, evaluates the result, and can refine the workflow from execution feedback. The main routes are task execution through the Darwinian workflow loop and optional goal decomposition through the planner; CSV evaluation provides a benchmark-oriented route. The README also documents reproducibility and audit trails. Edges reflect sampled imports/calls or explicit README claims; some internal evaluation and workspace details are unsampled, so those connections are omitted where evidence is insufficient.
Main entry points: dgm.py Darwin workflow loop · main.py CLI entry · onboard_cli.py Onboarding · workflow_factory.py Workflow factory
Components 18
User entry
- CLI entry
main.py - Onboarding
sources/cli/onboard_cli.py - Configuration
config.py
Planning and discovery
- Goal planner
sources/core/planner.py - MCP discovery
sources/core/tools_manager.py - LLM provider
sources/core/llm_provider.py
Workflow execution
- Darwin workflow loop
sources/core/dgm.py - Workflow orchestrator
sources/core/orchestrator.py - Workflow factory
sources/core/workflow_factory.py - Sandbox runner
sources/core/workflow_runner.py - Agent state
sources/modules/state_schema.py - Workflow records
sources/core/workflow_info.py
Evaluation and learning
- Workflow judge
sources/evaluation/evaluator.py - Scenario loader
sources/evaluation/scenario_loader.py - Benchmark mode
sources/evaluation/csv_mode.py
Other components
- Researcher
- Toolomics MCP
- ScienceAgentBench
Connections 24
- Researcher → CLI entry: starts
README.md - CLI entry → Onboarding: launches
main.py - CLI entry → Goal planner: routes goals
main.py - CLI entry → Darwin workflow loop: routes tasks
main.py - CLI entry → Benchmark mode: starts evaluation
main.py - Onboarding → MCP discovery: discovers tools
sources/cli/onboard_cli.py - Onboarding → Goal planner: launches goal
sources/cli/onboard_cli.py - Onboarding → Darwin workflow loop: launches task
sources/cli/onboard_cli.py - Configuration → MCP discovery: configures discovery
sources/core/tools_manager.py - MCP discovery → Toolomics MCP: discovers MCPs
README.md - Goal planner → Darwin workflow loop: executes steps
sources/core/planner.py - Goal planner → LLM provider: requests planning
sources/core/planner.py - Darwin workflow loop → Workflow orchestrator: runs workflow
sources/core/dgm.py - Darwin workflow loop → Workflow judge: scores execution
sources/core/dgm.py - Darwin workflow loop → Workflow records: tracks runs
sources/core/dgm.py - Workflow orchestrator → Workflow factory: crafts workflow
sources/core/orchestrator.py - Workflow orchestrator → Sandbox runner: runs sandbox
sources/core/orchestrator.py - Workflow factory → MCP discovery: uses MCP tools
sources/core/workflow_factory.py - Workflow factory → LLM provider: generates workflow
sources/core/workflow_factory.py - Workflow factory → Agent state: uses agent state
sources/core/workflow_factory.py - Workflow judge → Scenario loader: loads scenarios
sources/evaluation/evaluator.py - Benchmark mode → Darwin workflow loop: runs tasks
sources/evaluation/csv_mode.py - Benchmark mode → ScienceAgentBench: loads benchmark
sources/evaluation/csv_mode.py - Benchmark mode → Workflow judge: evaluates results