AstralDeep: bounded agents, useful interfaces

An open agent ecosystem

AstralDeep is an open ecosystem for agents that can do useful work.
We're building it around a simple goal: people should understand what an agent can do, even when it delegates.

Authority follows the task

An agent may discover a tool, ask another agent for help, or act on sensitive information.
Each step needs a clear identity and permission boundary.
Finding a capability should never grant permission to use it.

Bounded delegation with LETS

With LETS enforcement enabled, delegation has explicit limits.
A child lease receives only a bounded share of its parent's authority and resource budget, with an expiry.
The executor verifies and durably claims a receipt before the protected action.

Checks at the action boundary

AstralDeep's design keeps identity, tool policy, and owner isolation on the normal dispatch path.
Sensitive actions need the appropriate checks and confirmation.
Audit provenance helps explain who authorized an action and how authority reached it.

Generative UI with primitives

Useful agents also need useful interfaces.
A model can propose a form, table, or chart using structured UI primitives.
Code checks component types and references, bounds the layout, and escapes web content before rendering.
The model proposes. Code decides.

One contract across clients

AstralPrimitives defines the vocabulary.
AstralProjection renders and adapts it for web and native clients.
AstralDeep orchestrates the work, while AstralPlane owns durable state.
The goal is consistent meaning across devices, with components other developers can reuse.

Open protocols and reusable parts

We want these parts to connect through standard MCP, Agent2Agent, and other agent communication protocols.
That means versioned contracts, compatibility tests, and examples that work outside the full Astral stack.

Work that moves this forward

There is concrete work to do.
Help test delegation boundaries, improve protocol compatibility, or make voice survive interruptions and reconnects, especially on mobile.
Better primitive rendering, client parity, and independent examples can make the system useful to more developers.

A place to contribute

The bounty board turns these goals into tasks across five repositories.
Accepted, reviewed work earns recognition points on the leaderboard.
Join us at astraldeep.github.io.