How overmind-core/overmind works
Overmind is a platform for continuously improving specialized agents from production traces or other datasets. Users connect an agent, collect and organize data, define capability-level evaluations, run optimization or fine-tuning, then benchmark and serve models. The Console is one entry point; the REST API and MCP server expose platform workflows too. The README documents the product stages and capabilities, while sampled source confirms selected MCP, inference, and training internals; unsampled subsystems are represented at their documented directory level, with relationships limited to documented flows.
Main entry points: modal_vllm_worker.py vLLM serving · server.py MCP server · catalog.py MCP tool catalog · services Evaluation workflows
Components 32
Console
- Web console
frontend/src
Agent data
- Capability graph
overbae/models/capabilities.py - Trace observability
overbae/models/traces.py - Datasets
overbae/api/datasets.py
Evaluation and improvement
- Evaluation records
overbae/models/evaluation.py - Evaluation workflows
overbae/services - Optimizer experiments
overbae/models/optimizer.py
Model lifecycle
- Fine-tuning jobs
overbae/models/finetuning.py - SFT training
overbae/services/sft_assets/engine_unsloth.py - Model families
modal_shared/modelfam/spec.py - Model registry
overbae/core/model_registry.py - Deployment orchestration
overbae/services/deployment.py - Deployment reconciler
overbae/tasks/inference_controller.py - Inference client
overbae/services/inference_client.py - vLLM serving
overbae/modal/modal_vllm_worker.py - Context budgeting
modal_shared/context_budget.py - Train/serve stacks
modal_shared/stacks.py - Modal workers
overbae/modal - Model weights
API and integrations
- Projects and identity
overbae/models/iam.py - MCP server
overbae/services/mcp/server.py - MCP tool catalog
overbae/services/mcp/catalog.py - MCP resources
overbae/services/mcp/resources.py - Evaluation tools
overbae/services/mcp/tools_evaluations.py - Fine-tuning tools
overbae/services/mcp/tools_finetuning.py - Connector tools
overbae/services/mcp/tools_connectors.py - MCP OAuth
overbae/services/mcp/oauth_routes.py - REST API
overbae/api
Other components
- Agent builder
- Instrumented agent
- Coding agent
- OpenAI-compatible API
Connections 31
- Agent builder → Web console: uses
README.md - Web console → REST API: manages platform
- Instrumented agent → Trace observability: emits traces
README.md - Trace observability → Datasets: forms datasets
README.md - Agent builder → Capability graph: scans repository
README.md - Datasets → Evaluation workflows: evaluates
README.md - Capability graph → Evaluation workflows: defines metrics
README.md - Evaluation workflows → Evaluation records: records results
- Evaluation workflows → Optimizer experiments: scores experiments
README.md - Optimizer experiments → Agent builder: returns git diff
README.md - Datasets → Fine-tuning jobs: trains on data
README.md - Fine-tuning jobs → SFT training: runs fine-tuning
- SFT training → Model families: resolves family
overbae/services/sft_assets/engine_unsloth.py - SFT training → Model weights: produces weights
- Model families → Train/serve stacks: selects stack
modal_shared/stacks.py - Deployment reconciler → Deployment orchestration: reconciles deployments
overbae/tasks/inference_controller.py - Deployment orchestration → Modal workers: deploys models
README.md - Inference client → vLLM serving: sends routed requests
overbae/services/inference_client.py - vLLM serving → Context budgeting: reserves output tokens
overbae/modal/modal_vllm_worker.py - vLLM serving → Model families: resolves family
overbae/modal/modal_vllm_worker.py - vLLM serving → Model weights: loads weights
- vLLM serving → OpenAI-compatible API: serves inference
README.md - Agent builder → Coding agent: connects coding agent
README.md - Coding agent → MCP server: calls tools
README.md - MCP server → MCP tool catalog: dispatches tools
overbae/services/mcp/server.py - MCP server → MCP resources: lists resources
overbae/services/mcp/server.py - MCP tool catalog → Evaluation tools: registers tools
overbae/services/mcp/catalog.py - MCP tool catalog → Fine-tuning tools: registers tools
overbae/services/mcp/catalog.py - MCP tool catalog → Connector tools: registers tools
overbae/services/mcp/catalog.py - MCP server → MCP OAuth: authenticates clients
overbae/services/mcp/server.py - Coding agent → REST API: uses REST API
README.md