The project management system
built for AI agents

Loom is an MCP-native project management server. AI assistants connect via the Model Context Protocol and get full project management capabilities in under 3,000 tokens of context — with a web dashboard that keeps humans in the loop.

Connect your AI assistant

Works with Claude Code, Claude Desktop, Cursor, Windsurf, and any MCP-compatible client.

Why AI needs its own project management

Traditional project management tools were designed for humans clicking through web interfaces. When an AI assistant needs to plan work, track progress, or build a spec, it resorts to writing markdown files or dumping unstructured text. Loom gives AI agents the same rigour humans expect — typed schemas, versioned milestones, dependency graphs, and evidence-gated delivery — through native MCP integration.

Without Loom

  • AI writes TODO lists in markdown files
  • No structure — items are just strings
  • No links between tasks, specs, or decisions
  • No versioning — previous state is lost
  • Human has to read AI output to find status

With Loom

  • AI uses typed MCP tools with field validation
  • 10 item kinds with schemas and status machines
  • Dependency graphs: blocks, implements, tests
  • Versioned scopes with closeout and promotion
  • Human opens a dashboard and sees everything

One system. Every workflow.

Loom uses typed artifacts to handle goals, specifications, components, and invariants in a single unified model. Goals track priorities and due dates. Scopes version your work into milestones. Mix and match within the same project.

Goals & Operations

For the work that keeps a company running. Goal artifacts with priority levels, due dates, and automatic progress tracking. Tasks show status bullets, blocked detection, tags, and expandable detail panels. Collapsible sections keep long goal lists scannable.

  • Goals with critical / high / medium / low priority sorting
  • Blocked item detection across dependency links
  • Markdown content attachments on any item
  • Progress counters and due-date-aware ordering

Specifications & Architecture

For building software with structure. Versioned scopes organise work into plannable milestones. Artifacts group related items — components, invariants, decisions, scenarios. Close a scope and implemented work promotes to the canonical project spec.

  • Component matrices with capabilities and dependencies
  • Invariants with rule, rationale, and enforcement fields
  • Scope versioning: v0.1.0 → v0.2.0 → closeout
  • Artifact-level progress tracking per scope

What your AI actually gets

Full project management — projects, scopes, goals, items, links, tags, search, activity, export — consolidated into 14 tools that fit in ~2,700 tokens. Most MCP servers burn 500+ tokens per tool. Loom uses a consolidated design where one tool handles all mutations and another handles all reads, keeping the context footprint minimal while covering the full domain.

Typed Items

10 built-in item kinds: task, component, invariant, decision, scenario, question, narrative, note, and more. Each kind has its own field schema and status machine. The AI knows the shape of every item before it creates one.

Links & Dependencies

Polymorphic, typed links: depends_on, blocks, implements, tests, enforces, justifies. The AI navigates a real dependency graph — finding blocked items, tracing requirements to implementations, following test coverage.

Content Storage

Attach markdown documents to any item. Content-addressable storage with SHA-256 deduplication. Build a project knowledge base the AI can reference across sessions without losing context.

Evidence Gating

Items don't just get marked done. Verification runs record commit SHAs, test results, and review approvals. Quality is structural — the system enforces it, not a process document.

Tags & Search

Tag any entity across projects. Full-text search from the dashboard with live results. The AI uses tags to organise cross-cutting concerns; humans use them to filter and navigate.

Human Dashboard

Everything the AI creates is visible in a real web interface. Expandable item rows, collapsible sections, activity feeds with direct links, component grids, and invariant cards. Not an afterthought — a first-class UI.

How it works

1

Connect

Copy the MCP server URL into your AI client. Authenticate via OAuth 2.1. Your AI gets scoped access to your workspace in seconds.

2

Build

Your AI creates projects, goals, scopes, and items using structured MCP tools. Every action is validated, typed, and automatically tracked.

3

Review

Open the web dashboard to see what your AI built. Expand items for detail, collapse sections for overview, click through links. Full visibility.

Built for the Model Context Protocol

Loom is a purpose-built MCP server, not a traditional SaaS with an API adapter. The entire data model — projects, scopes, artifacts, items, links, tags, and content — is exposed through native MCP tools with JSON Schema input validation and structured output.

MCP Integration

Protocol
Model Context Protocol (MCP) over streamable HTTP
Authentication
OAuth 2.1 with PKCE, scoped access tokens
Tools
14 consolidated tools covering the full project management domain
Context cost
~2,700 tokens for the entire toolset (1.3% of a 200k context window)
Design
Consolidated: one write tool for all mutations, one get/list for all reads
Schemas
Every tool has typed JSON Schema inputs and structured JSON outputs
Resources
MCP resources for project overviews and scope summaries
Prompts
Guided workflows for common operations (new project, scope planning)

Data Model

Item kinds
task, component, invariant, decision, scenario, question, narrative, note, epic, bug
Link types
depends_on, blocks, implements, tests, enforces, justifies, relates_to
Priorities
critical, high, medium, low — with automatic sort ordering
Status machines
Per-kind status flows (planned → in_progress → implemented)
Content
Markdown with content-addressable SHA-256 deduplication
Versioning
Semantic scope versions with closeout and spec promotion