AXME Code

Stop re-explaining your architecture on session 47. Stop losing memory between session handoffs. Stop hoping the agent won't rungit push --forceto main. AXME Code remembers what happened, enforces your architectural decisions, continues where the last session stopped, and blocks dangerous commands before they execute — so you can focus on building.

You keep using Claude Code exactly as before. AXME Code works transparently in the background.

Session 1: "We use FastAPI, not Flask. Deploy only via GitHub Actions. Never push to main directly."

Session 2: "Like I said yesterday, we use FastAPI..."

Session 7: "For the third time this week, we use FastAPI..."

Session 47:gives up, mass-pastes 200 lines into CLAUDE.md

Session 1: Agent learns your stack, saves decisions.

Session 2: Agent callsaxme_context→ already knows FastAPI, deploy rules, what happened yesterday.

Session 47: Agent has your full project history: 30 decisions, 15 memories, safety rules, and a handoff from session 46.

Agent runsgit push --forceto main. Your Friday is ruined.

Hook intercepts the commandbefore executionand blocks it. Not a prompt — hard enforcement at the harness level.

Agent says "Done!" — but tests don't pass, half the code is stubbed, and the deploy is broken.

Decisions enforce verification requirements: agent must run tests and show proof before reporting completion.

AXME Code supports three IDE paths today, ranked by lowest install friction:

Option 0: Cursor extension (1-click install — recommended for Cursor users)

ForCursor 0.42+users — install theAXME Codeextension from the Extensions panel (Open VSX). The extension bundles the binary, registers the MCP server programmatically (no manual Enable click), installs user-level safety hooks at~/.cursor/hooks.json(apply to every project on your machine), and offers a one-click "Run setup" notification the first time you open a project without.axme-code/.

Cursor → Extensions → search "AXME Code" → Install

On first activation a modal asks for an LLM credential for the session auditor: paste an Anthropic API key, a Cursor SDK key (cursor.com → Integrations), or skip the auditor. IfclaudeCLI is logged in (claude login), the extension auto-uses your Claude subscription — no paste needed.

Option 1: Claude Code plugin (recommended for Claude Code users)

/plugin marketplace add anthropics/claude-plugins-community /plugin install axme-code@claude-community

claude plugin marketplace add anthropics/claude-plugins-community claude plugin install axme-code@claude-community

The plugin ships with the MCP server, safety hooks, and CLI bundled together; no separate binary to install. On first use in a project, just ask the agent to callaxme_context— the plugin auto-initializes the knowledge base on that session.

Install the CLI system-wide (useful if you want to runaxme-codeoutside Claude Code, e.g. for scripting).

curl -fsSL https://raw.githubusercontent.com/AxmeAI/axme-code/main/install.sh | bash

Installs to~/.local/bin/axme-code. Requires Node.js 20+ on PATH (the binary is a single-file Node bundle; the installer checks for it). Supports x64 and ARM64.

irm https://raw.githubusercontent.com/AxmeAI/axme-code/main/install.ps1 | iex

Installs to%LOCALAPPDATA%\Programs\axme-codeand adds it to your User PATH. Requires Node.js 20+ on PATH. Supports x64 and ARM64.

Windows via WSL2:if you already live in WSL2, use the Linux install one-liner inside your distro. Install Claude Code and axme-codeinsidethe WSL distro, not on the Windows host.

cd your-project # or workspace root for multi-repo axme-code setup claude # that's it — use Claude Code as usual

- Scans your projectand builds the knowledge base — oracle (stack, structure, patterns, glossary), extracts decisions, memories, and safety rules from your code, configs, CLAUDE.md, and session history
- Installs safety hooksthat intercept dangerous commands before execution
- Configures the MCP serverin Claude Code settings (.mcp.json)

After setup, every Claude Code session automatically loads the full knowledge base. No config, no manual steps.

Your agent starts every session with full context: stack, decisions, patterns, glossary, and a handoff from the previous session. No more re-explaining your architecture on session 47.

Hooks intercept tool callsbefore execution— not prompts. Even if the agent hallucinates a reason to runrm -rf /, the hook blocks it. This is hard enforcement at the Claude Code harness level, not a suggestion in a system prompt.

- git push --force,git reset --hard, direct push tomain/master
- rm -rf /,chmod 777,curl | sh
- npm publish,git tag,gh release create
- Writing to.env,.pem,.keyfiles

You can add your own custom rules viaaxme_update_safetyor by editing.axme-code/safety/rules.yamldirectly.

The agent saves discoveries during work via MCP tools. At session close, a structured checklist ensures nothing is missed. If you just close the window — a background auditor extracts memories, decisions, and safety rules from the full session transcript.

Each repo gets its own knowledge base (.axme-code/). Workspace-level rules apply to all repos. Repo-specific rules stay scoped. The agent sees merged context — workspace safety floor + repo-specific decisions.

Supports 14 workspace formats: VS Code multi-root, pnpm/npm/yarn workspaces, Nx, Gradle, Maven, Rush, git submodules, and more.

CLAUDE.md is great for simple projects with a few rules. But it doesn't scale:

AXME Code complements CLAUDE.md — it reads your existing CLAUDE.md during setup and extracts decisions and rules from it.

The MCP server entry is identical in every client:

AXME uses~10× fewer tokens per correct answerthan Mastra at competitive accuracy. The memory system runs only 2 LLM calls per question (reader + judge) — competitors run dozens (Observer per turn, Reflector periodically, graph construction, fact extraction).

Tip: You can save at any time — just tell the agent "remember this" or "save this as a decision". You don't have to wait for session close.

All data lives in.axme-code/in your project root (gitignored automatically):

.axme-code/ oracle/ # stack.md, structure.md, patterns.md, glossary.md decisions/ # D-001-slug.md ... D-NNN-slug.md (with enforce levels) memory/ feedback/ # Learned mistakes and corrections patterns/ # Validated successful approaches safety/ rules.yaml # git + bash + filesystem guardrails backlog/ # B-001-slug.md ... persistent cross-session tasks sessions/ # Per-session meta.json (tracking, agentClosed flag) plans/ handoff-<id>.md # Per-session handoff (last 5 kept) worklog.jsonl # Structured event log worklog.md # Narrative session summaries config.yaml # Model settings, presets

Human-readable markdown and YAML. No database, no external dependencies.

1. MCP Server (persistent, runs while VS Code is open)

Provides tools for the agent to read and write the knowledge base. All writes go through MCP server code (atomicWrite, correct append) — the agent never writes storage files directly.

pre-tool-use: Checks every Bash command, git operation, and file access against safety rules. Blocks violations before execution. Also creates/recovers session tracking.

post-tool-use: Records which files the agent changed (for audit trail).

3. Background Auditor (runs after session close)

A detached process that reads the session transcript and catches anything the agent forgot to save. Two modes:

- Full extraction— when the agent crashed or user closed without formal close
- Verify-only— when the agent completed the close checklist (lighter, cheaper)