Capability Bootstrap

Capability Bootstrap is an open spec and reference runtime for one core agent behavior:

when an agent realizes it lacks a capability, it should be able to discover, acquire, validate, and reuse that capability with minimal human involvement.

This project is not a skill marketplace, not an MCP directory, and not a one-off installer. It is a reusable capability layer that other agent products can adopt.

Core idea

Most agent systems still assume that humans are responsible for expanding agent capability. When an agent needs a new MCP server, skill, tool adapter, or integration, the user is expected to:

• search manually
• read docs manually
• edit config manually
• install dependencies manually
• test and debug manually

Capability Bootstrap exists to remove that burden.

Humans should not need to learn how to provision agent capability.

What this project contains

• an open protocol for capability acquisition
• a policy layer for safe automation
• a state machine for the capability lifecycle
• a tiny reference runtime for end-to-end bootstrap flow
• a pluggable adapter interface for capability sources and preferences

Reference runtime

The runnable implementation currently models two adapter categories:

• source adapters: discover capability candidates from one source
• preference adapters: annotate or boost candidates after discovery

Current built-in adapters:

• skill-source-adapter
• mcp-source-adapter
• memory-reuse-adapter

The end-to-end flow is:

1. detect a capability gap
2. ask source adapters for candidates
3. apply preference adapters
4. rank low-risk candidates
5. apply policy
6. install or connect a candidate
7. validate it
8. remember it for reuse

Quick start

cd D:\AI Pro\capability-bootstrap
node src/demo.js browser-control
node src/demo.js github-workflow
node src/demo.js reset-memory

Current demo scope

The demo intentionally stays small. It simulates:

• a missing browser-control capability
• a missing github-workflow capability
• remembered reuse of previously validated candidates
• candidate aggregation across skill and MCP sources

Project structure

• README.md: project entry point
• SPEC.md: public protocol definition
• POLICY.md: automation and safety boundaries
• STATE-MACHINE.md: lifecycle definition
• docs/vision.md: product rationale
• docs/mvp.md: first implementation boundary
• docs/next-steps.md: suggested implementation path
• docs/adapter-interface.md: pluggable adapter contract
• src/adapters/: source and preference adapters
• src/catalogs/: tiny in-repo candidate catalogs for the demo
• src/: reference runtime
• examples/: concrete bootstrap scenarios
• schemas/: machine-readable schema experiments
• data/: runtime memory output during demos

Initial design principles

• human-out-of-the-loop by default
• installed reality first
• discover only when necessary
• install only when justified
• verify before trust
• interrupt humans only for risk
• remember what was learned

Status

This repo now contains both the public design skeleton and a tiny runnable reference runtime. The next step is to replace demo catalogs with real source adapters without losing the safety boundary.