Skip to content

ml-rust/axp

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

31 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

AXP — Agent Execution Protocol

An open protocol that gives AI agents a secure, sandboxed place to run code and use tools — with lean context usage. AXP works alongside MCP.

⚠️ Early draft. AXP is in active design. The protocol and these documents will change. Feedback, discussion, and early implementations are very welcome.


Why AXP

When an AI agent needs to do real work — run a command, edit files, call a tool, start a long-running process — two things matter a lot: safety and context cost. AXP focuses on exactly that:

  • Safe execution. Each agent gets an isolated workspace. Every action is capability-scoped, sandboxed at the OS level, and audited. The host is not exposed.
  • Lean context. Tool discovery is lazy: the agent sees a compact index of what's available and loads full detail only for the tool it actually uses, so context stays small even with large catalogs.
  • Reliable long-running work. Jobs stream their output and can be detached from and reattached to, so work survives disconnects.
  • Works with what you have. AXP speaks MCP in both directions — mount existing MCP servers as capabilities, or expose AXP to MCP clients.

MCP connects agents to tools; AXP focuses on giving agents a safe place to run them. Many setups will use both.

How it fits together

Agent ──► AXP Session ──► isolated Workspace
                      ├─► Capabilities  (explicit, least-privilege, attenuable)
                      ├─► Sandbox       (OS-level: Landlock/seccomp · Seatbelt · AppContainer)
                      ├─► Jobs          (streaming logs, resumable, reattachable)
                      └─► Providers     (native · code-mode · MCP servers · skills)

A quick taste of the flow (illustrative):

session.open  { workspace, sandbox_tier, capabilities[] }   → { session_id }
axp.index     { session_id }                                → [{name, desc}, ]   // cheap, full catalog
axp.describe  { name }                                      → { signature, schema } // detail on demand
job.start     { command | code }                            → { job_id }          // streams logs, resumable

See docs/AXP-ARCHITECTURE.md for the full design.

Cross-platform

AXP is implemented in Rust and ships as a single static binary. It targets Linux, macOS, and Windows on x86_64 and arm64. Sandbox enforcement uses each platform's native mechanism, and a session always declares its enforcement tier so clients know the guarantee they're getting.

Status

Area State
Architecture draft In progress — see docs/
Runtime (Rust) Early development
MCP bridge Planned
SDKs (TypeScript / Python / Go) Planned
Conformance suite Planned

Contributing

AXP is developed in the open and welcomes contributions — see CONTRIBUTING.md. Design changes go through a public RFC process so the reasoning stays visible and reviewable.

License

Licensed under the Apache License 2.0. See LICENSE.

About

An open protocol that gives AI agents a secure, sandboxed place to run code and use tools — with lean context usage. AXP works alongside MCP

Resources

License

Code of conduct

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages