vzroslieigry/maxagent

What it does

maxagent is an experience layer for coding agents such as Claude Code, Codex, Antigravity and OpenClaw. Instead of each session rediscovering the same solution, agents read and write short checked recipes ("atoms") in one shared workspace, plus project state files that record where work stopped. Hooks injected into each agent surface the relevant atom on a user message, before a command, or after an error.

The repo ships two packages: the main maxagent CLI (rules, workspace, experience engine, per-agent hooks, setup wizard) and maxagent-bridge, an MCP server that drives the user's own Chrome profile through a small extension, with per-agent approval, a site blocklist, an action log and a stop button.

Quick start

Requires Node.js 18 or newer. Run the wizard:

npx maxagent init

The wizard creates a workspace at ~/maxagent by default (AGENTS.md, experience/, projects/, sync.md, inbox/), inserts a marked rules block and hooks into every supported agent it finds, and backs up files before editing. Extras are optional and gated by a single question:

• Browser bridge — npx maxagent bridge; load the Chrome extension unpacked via chrome://extensions (Developer mode → Load unpacked).
• Meaning search — npx maxagent vectors downloads the multilingual-e5-small model and Python libraries (~270 MB) into ~/.maxagent/vector; without Python, search falls back to word matching.
• Shared experience packs — npx maxagent experience add all.

Afterwards npx maxagent doctor reports what is installed and healthy. Settings can be changed with npx maxagent setup, and npx maxagent uninstall removes agent blocks and hooks while keeping the workspace. The package is also on npm and supports a one-phrase agent-driven install.

Working with experience

Your own atoms come from experience/ and are surfaced first; they can stop a command known to fail. Shared packs are reference-only: shown only when your shelf is empty, with a higher threshold, at most one compact atom per message, and they never stop a command. npx maxagent experience adopt "<id>" copies a useful shared atom into your own shelf. Packs are single JSON files on GitHub Releases, listed in packs/catalog.json with sha256 and size; the client checks the checksum, filenames and text (download-and-run commands, rule rewrites, secret exfiltration, destructive commands, hidden characters) and rejects the whole pack on any hit.

Privacy and constraints

Workspace files belong to the user; updates do not overwrite them. Network requests are plain GETs of public files (npm registry, a news.json note feed, and, only after consent, GitHub Releases and Hugging Face model files). Setting updates: false disables all requests; a daily background update check announces new versions once per session. The browser bridge uses Chrome Native Messaging and a local pipe/socket, with no open port. The extension must currently be loaded unpacked. Rule translations are agent-made and not yet checked by native speakers. Licensed under MIT.

HotRepo's take

On the JavaScript new-projects list for 2026-11-10, maxagent recorded 112 stars and 9 new stars over the measured window from 22:15 to 23:45 UTC, with no all-language daily rank captured. That is a very early sample for a project created two days earlier, so the numbers show initial attention rather than sustained adoption. The design choice worth watching is the split between writable personal atoms and checked, checksummed shared packs, which is the part that determines whether experience accumulates usefully or becomes noise.