A local, file-backed runtime for autonomous software delivery, with a Python CLI, a Textual TUI, separate execution and research planes, durable state, diagnostics, and restart-safe control.
The public runtime ships as a Python engine with a CLI, a Textual TUI, file-backed operational truth, frozen per-run plans, and restart-safe operator control.
Public runtime control flow
Preview the public runtime control flow here, then open the full architecture view to inspect package surface, workspace setup, control plane, shared engine, execution, research, operational truth, and publish surfaces.
These products are not all solving the same problem. Some are managed autonomous engineers. Some are terminal agents. Some are IDE products. Some are open platforms. Some are general-purpose assistants. Millrace is a runtime-first system for autonomous software delivery. The comparisons below are about category, control plane, and operator trust surface.
Model-agnostic runtime. Runner and model choice live in configuration, not in the control plane.
Devin is a managed autonomous engineering product built around Ask and Agent sessions, with Knowledge, Playbooks, session tools, and APIs. Millrace differs by emphasizing a local, file-backed runtime with explicit operator-visible control state.
Devin
What it is
Managed autonomous engineering product.
Runtime surface
Ask mode, Agent mode, Knowledge, Playbooks, session tools, and APIs.
Control model
Managed product surface with autonomy inside the session environment.
Strength
Best when you want fast adoption and a polished managed experience.
Millrace
What it is
Local, file-backed runtime for autonomous software delivery.
Runtime surface
Python CLI, Textual TUI, execution plane, research plane, durable state, and diagnostics.
Control model
Plans, queues, provenance, and daemon controls stay operator-visible and restart-safe.
Strength
Best when you want local runtime authority and inspectable control state.
The core difference: Devin packages agentic execution as a managed product. Millrace packages it as a local runtime you can inspect and steer directly.
Claude Code is no longer just a single terminal coding agent. Its public surface now includes subagents, permission modes, and preview autonomy features. Millrace differs by keeping the runtime deterministic and operator-legible instead of using the agent product as the control plane.
Claude Code
What it is
Terminal-native coding agent.
Current surface
Subagents, permission modes, and expanding autonomy previews.
Control model
The product decides execution flow inside the session and manages long work through agent-side context handling.
Strength
Best when you want a powerful terminal agent inside Anthropic's product surface.
Millrace
What it is
Runtime-first autonomous delivery system.
Current surface
File-backed plans, queues, diagnostics, execution plane, and research plane.
Control model
Runner choice lives in configuration. The control plane lives in files and state, not in an agent session.
Strength
Best when you want architecture-level governance over long-horizon work.
The core difference: Claude Code is a strong terminal agent. Millrace is a runtime that can use agents without making them the source of truth.
Cursor is no longer just an AI IDE copilot. It now includes background agents, a Background Agents API, web and mobile entry points, and Bugbot for PR review. Millrace differs by living outside the editor as a runtime-first autonomous delivery system.
Cursor
What it is
Agentic IDE product.
Current surface
Background Agents, API, web/mobile entry points, and Bugbot.
Control model
Editor-centric workflow with growing remote execution surfaces.
Strength
Best when the developer still lives inside the IDE and wants faster delivery loops.
Millrace
What it is
Runtime-first autonomous delivery system.
Current surface
CLI, TUI, file-backed state, frozen per-run plans, and diagnostics.
Control model
Editor optional. Runtime state stays outside the IDE.
Strength
Best when you want unattended or semi-unattended long-horizon execution with auditable control.
The core difference: Cursor extends the IDE into agentic execution. Millrace moves the control plane outside the IDE entirely.
Windsurf is no longer just Cascade in an editor. It now includes reusable Workflows and App Deploys alongside its core agentic IDE experience. Millrace differs by being a standalone runtime rather than an editor-led workflow product.
Windsurf
What it is
Agentic IDE product.
Current surface
Cascade, Workflows, and App Deploys.
Control model
Editor-led workflow automation with reusable sequences.
Strength
Best when you want coding, workflow reuse, and deployment helpers inside the IDE experience.
Millrace
What it is
Standalone autonomous delivery runtime.
Current surface
Explicit execution and research planes, queue and status surfaces, frozen plans, and diagnostics.
Control model
File-backed control plane, not an editor workflow layer.
Strength
Best when you want one runtime to carry work forward outside the IDE.
The core difference: Windsurf keeps the developer at the center of an editor-led agent workflow. Millrace keeps the runtime at the center of an autonomous delivery loop.
OpenHands is better described as an open agent platform and SDK stack with local and remote workspaces, tools, and agent-server deployment modes. Millrace differs by being a narrower, more opinionated runtime for autonomous delivery.
OpenHands
What it is
Open agent platform and SDK stack.
Deployment model
Local workspaces, remote or sandboxed workspaces, tools, and agent-server modes.
Control model
Composable and self-hostable, with more moving abstractions.
Strength
Best when you want open infrastructure for building and running agents.
Millrace
What it is
Opinionated autonomous delivery runtime.
Deployment model
Local, file-backed runtime with explicit queue, status, plan, and provenance surfaces.
Control model
Narrower surface, stronger runtime discipline, and operator-visible state.
Strength
Best when you want a dedicated software-delivery control plane.
The core difference: OpenHands is an open platform for composing agents. Millrace is a tighter runtime for carrying software delivery forward.
Factory is an AI-native software development platform built around Droids that work across IDE, web, CLI, Slack, and project-management surfaces. Millrace differs by being a local, file-backed runtime rather than a multi-surface agent product.
Factory AI
What it is
AI-native software development platform.
Current surface
Droids across IDE, web, CLI, Slack, and project manager workflows.
Control model
Auto-run modes, risk levels, and workflow integration live inside the product environment.
Strength
Best when you want agent-native software development inside the tools your team already uses.
Millrace
What it is
Local, file-backed runtime for autonomous software delivery.
Current surface
Python CLI, Textual TUI, execution plane, research plane, and durable diagnostics.
Control model
Frozen per-run plans, operator-visible provenance, and restart-safe control live in the runtime itself.
Strength
Best when you want one inspectable control plane rather than many product surfaces.
The core difference: Factory spreads agents across the software stack. Millrace concentrates control into a single runtime that owns delivery state directly.
Attractor is presented as part of StrongDM's Software Factory: a non-interactive coding agent structured as a graph of phases. Millrace differs by favoring a more direct, file-backed runtime control plane over a graph-centric product story.
Attractor
What it is
Non-interactive coding agent structured as a graph of phases.
Execution model
Specs and scenarios drive convergence through the broader Software Factory system.
Control model
Graph-centric product story with phase transitions as the core abstraction.
Strength
Best when fully specified work sits inside a broader software-factory stack.
Millrace
What it is
File-backed autonomous delivery runtime.
Execution model
Explicit execution and research planes with frozen per-run plans and visible runtime state.
Control model
Direct, inspectable runtime surfaces rather than graph-centric abstractions.
Strength
Best when you want a runtime you can inspect and steer directly.
The core difference: Attractor packages autonomy as part of a broader factory stack. Millrace packages it as a direct runtime with visible delivery state.
GSD is a Claude Code-native workflow system focused on context engineering, spec-driven planning, and fresh subagent execution. Millrace differs by being its own runtime rather than a workflow layer on top of Claude Code.
GSD
What it is
Claude Code-native workflow layer.
Current surface
Meta-prompting, spec-driven phases, roadmap and phase commands, and fresh subagent execution.
Control model
Adds structure around Claude Code, but execution still depends on the agent product underneath.
Strength
Best when you already live inside Claude Code and want more structure without a separate runtime.
Millrace
What it is
Runtime-first autonomous delivery system.
Current surface
Own control state, explicit runtime surfaces, file-backed plans, and diagnostics.
Control model
Agents run inside the runtime. They do not define the runtime.
Strength
Best when you want orchestration authority to live outside the agent product.
The core difference: GSD improves agent workflow from within Claude Code. Millrace replaces the agent-led loop with its own runtime control plane.
BMAD is better described now as a broader AI-driven development ecosystem with modular workflows, specialized agents, scale-adaptive planning tracks, and customization infrastructure. Millrace differs by optimizing for a leaner autonomous runtime.
BMAD
What it is
AI-driven development ecosystem.
Current surface
Specialized agents, workflows, modules, installers, and planning tracks.
Control model
Strong guided process structure and artifact generation.
Strength
Best when you want modular, human-steered development process scaffolding.
Millrace
What it is
Leaner autonomous runtime.
Current surface
File-backed truth, explicit control-plane surfaces, and long-horizon execution loops.
Control model
Less ecosystem breadth, more runtime specificity.
Strength
Best when you want runtime discipline and unattended delivery.
The core difference: BMAD expands the process ecosystem around the operator. Millrace narrows the surface to a runtime that can carry delivery on its own.
Manus is a cloud general-purpose autonomous agent platform. It spans a much broader task surface than software delivery alone. Millrace differs by being purpose-built for autonomous software development rather than general agent breadth.
Manus
What it is
Cloud general-purpose autonomous agent.
Task surface
Broad async work across research, operations, and mixed workflows.
Control model
Generalist agent product with broad task coverage.
Strength
Best when you want one agent across many categories of work.
Millrace
What it is
Specialized autonomous software-delivery runtime.
Task surface
Software building, validation, recovery, and operator governance.
Control model
Purpose-built delivery state control rather than general cross-domain breadth.
Strength
Best when you want software-building depth over general utility.
The core difference: Manus optimizes for breadth across task categories. Millrace optimizes for depth in software delivery.
OpenClaw is now much more than a viral local agent demo. It is a local-first personal AI assistant ecosystem with workspaces, skills, sessions, channels, and a growing plugin surface. Millrace differs by being purpose-built for autonomous software delivery rather than general personal-agent breadth.
OpenClaw
What it is
Local-first personal AI assistant ecosystem.
Current surface
Workspaces, skills, sessions, channels, and broader ecosystem tooling.
Control model
Broad local-agent surface that should be treated as high-trust software and isolated appropriately.
Strength
Best when you want a general local agent across personal workflows and device surfaces.
Spec to execution to validation to recovery inside a dedicated control plane.
Control model
Narrower scope, stronger delivery-state discipline, and operator-visible provenance.
Strength
Best when you want dedicated software-building depth rather than general agent breadth.
The core difference: OpenClaw maximizes breadth across local agent workflows. Millrace maximizes depth inside a dedicated software-delivery runtime.
Built by Tim Osterhus — from fitness coaching to autonomous software, with no engineering background. If one person with the right framework can compete with billion-dollar labs, what does that say about the framework?
Follow the runtime.
Get notified when major runtime releases, public demos, and compiler milestones land.
No spam. Just major runtime releases, public demos, and compiler milestones.
