Skip to content
中文
Sensing, Memory, and Action on a plate, with HALO's ring above themSensing, Memory, and Action on a plate, with HALO's ring above them

HALO

Plug HALO into a personal agent through MCP. The agent logs each step of its loop, and HALO shows which channels carry information, how far its evidence for self-improvement goes, and which level the trace supports.
Any MCP client
{
"mcpServers": {
"halo": { "command": "halo", "args": ["serve"] }
}
}

A plugin, not a fork

HALO is an MCP server. Hosts that take a custom MCP server, such as ChatGPT dots, Grok Bot, Claude Code, and Codex, can use it without changing their code. How it connects.

Channels from links

Every step the agent logs names the earlier steps it used. A Memory step used by an Action step is the channel M→A. Event model.

Evidence for self-improvement

HALO follows an action to its outcome, the revision the outcome led to, and the later action that used the revision. Analysis.

Levels as lower bounds

The report places the trace at a loop-closure level from L0 to L5, and says what the next level would need. Levels.

You want to Start with
See what HALO reports Quickstart
Understand the plugin Plugin architecture, then Event model
Connect an agent The connection protocol, then a case such as ChatGPT dots
Connect a local agent Claude Code, Codex, and other local hosts
Look up a tool or route Plugin reference
Review published systems Read mode and Run mode