Todos

Todos is a task-driven workspace that lets humans and AI agents work together on real projects. A "Chief" agent takes a one-sentence goal, splits it into discrete todos, assigns each to the right specialist agent, and tracks progress — reporting back automatically as stages move forward. The product is built for solo builders and small teams of up to five people, and it runs builds on machines you own using model API keys you already have.

What It Is

Todos is a multi-agent project management and execution platform. Unlike chat-based coding assistants that hand back a result, Todos maintains a persistent board of todos, each of which is a live conversation with an assigned agent. The Chief agent acts as a team lead: it plans, delegates, and monitors without writing code itself. Specialist agents — frontend engineers, backend engineers, QA, DevOps, copywriters, and more — each carry their own model assignment, thinking level, and skill set. The result is a structured workflow that mirrors how a small human team operates, but with AI agents filling the specialist roles.

How the Agent Workflow Runs

Every build in Todos follows a layered execution model:

• Goal → Todos: You give Chief a sentence; it proposes a list of todos with assignments and dependency order (parallel where possible, sequential where required).
• Plan → Build: Each todo runs in two stages. A planner agent drafts an approach first; after you confirm, a builder agent implements it. The plan, diff, and test results all stay on one thread.
• AI Review: A separate reviewer agent can critique the plan or the change before it reaches you, with feedback returned to the thread and the original agent revising automatically.
• Merge: Approved changes open a PR and mirror to your local worktree in about a second.

Scheduled reruns let routine todos — daily reports, regression checks, content pipelines — fire on an hourly, daily, or weekly cadence without manual intervention.

Model Flexibility and Cost Transparency

Todos supports over 700 models across Anthropic, OpenAI, Google, DeepSeek, local models via Ollama, and any OpenAI-compatible endpoint. Each agent slot independently picks its provider, model, and thinking level, so demanding tasks can run high-capability models while simpler tasks use lighter, cheaper ones. ChatGPT/Codex and GitHub Copilot subscriptions can also be connected with a one-time sign-in, bypassing the need for a separate API key.

The platform itemizes token usage and estimated cost per model for every build, letting teams retune model assignments based on actual spend. Todos states that API keys stay on your machines and inference is never metered or marked up — you pay providers directly.

Deployment Model: Your Machines, Your Keys

Execution happens on hardware you own. The tds start CLI command registers a computer as an executor; builds run there in isolated git worktrees, and PRs go straight to GitHub. The Todos server coordinates projects, conversations, and reviews, but no GPU is required — any machine running Node.js and git qualifies. This architecture means build speed and parallelism scale with the machines you enroll, not with a vendor sandbox.

An MCP server ships with the product, allowing Claude Code, Cursor, VS Code, or any MCP client to read the board, create todos, and start builds using a single API key with configurable tool grants.

Agent Memory and Skills

Agents accumulate durable memory across rounds — project conventions, preferred formats, lessons from past failures — and reuse it automatically on future tasks. Skills are reusable SKILL.md playbooks that can be shared across the team, imported from any GitHub repo in one click, or pinned to an agent as defaults. High-impact capabilities like branch merging and remote shell access are off by default and granted per agent, giving fine-grained control over what each agent can do.

Platform and Access

Todos is accessible as a web app and installs as a PWA for mobile use, supporting voice input for capturing ideas, confirming plans, and reviewing results from a phone. Push notifications arrive when a decision is needed. The CLI (tds) is the only local installation required to bring a machine online as an executor.