Two services. Card 1 (Safe Edit Loop) runs invisibly via PreToolUse + PostToolUse hooks — deterministic in Claude Code, prompt-driven in Cursor / Aider / Cline. Card 2 (Red Team Audit) is one slash command before major merges. Same engine, two cadences, zero re-configuration when you switch between agents.
Claude Code refactors a function and breaks 7 callers nobody noticed. Cursor invents user.send_email() when the actual method is user.email(). The agent declares "Done!" — file contains throw new Error("not implemented") on three branches. You've paid $200/mo for Claude Max or Cursor Ultra for the privilege.
The fix isn't more prompting. It's removing the LLM's option to skip the check. ArgosBrain installs hooks that fire deterministic in Claude Code (the agent literally can't refuse) and auto-applied rules in Cursor / Aider / Cline (the agent is strongly guided, ~90-95% compliance on Opus 4.7).
Before every Edit / Write / MultiEdit, a PreToolUse hook calls mcp__argos__preflight with the symbol being modified. Returns existence + caller count + production/test split + risk verdict (LOW / MEDIUM / HIGH). Injected as additionalContext for the agent — it decides with real data, not vibes.
After every edit, a PostToolUse hook scans for stub patterns across 50+ languages (Rust todo!(), Python NotImplementedError, JS not-implemented throws, etc.). Hit → agent obligated to fix before claiming "done".
Compatibility tiers: Claude Code = 100% deterministic (hooks intercept before the LLM commits). Cursor = ~90-95% via .cursor/rules/argosbrain.mdc with alwaysApply: true. Aider / Cline / Codex CLI = ~80-90% via CONVENTIONS.md / .clinerules / Codex prompts. Full compatibility table →
/argos-security. Wait 30-50 min. Get a kill-chain report.Nine attacker perspectives run in parallel — Recon, Web/API, Cloud, AI/LLM, Supply-chain, Build & Release, Forgotten Attack Surface (dead code as attack surface), Surface Drift Watch (new endpoints since last release), Privilege Boundary Leaks (admin reachable from public). Composed into MITRE ATT&CK kill chains, ranked by (impact × confidence) / cost-to-exploit.
Static control-flow only — PoC pseudocode for your team's verification, no live exploits. Run weekly or before major merges. See the nine perspectives →
argosbrain init. Both cards wired automatically.curl -fsSL https://argosbrain.com/install.sh | sh cd ~/my-project argosbrain init --install-config .
Drops 5 hooks in ~/.claude/hooks/, registers them per-project in <project>/.claude/settings.json, writes .cursor/rules/argosbrain.mdc + .clinerules + CONVENTIONS.md + CLAUDE.md, installs /argos-security + /argos-doctor slash commands. Restart your agent. Done.
The agent now has 15 MCP tools visible in its catalog (lean by default; ARGOSBRAIN_EXPOSE=full opens all ~59 for power users). Card 1 hooks fire automatically on edits. Card 2 is one slash command away.
One engine, every agent. Switch tools without losing context — Card 2 invocation is identical (/argos-security), Card 1 enforcement adjusts to what each runtime supports. Memory lives on disk, not inside the editor.
Ingestion, storage, retrieval, hooks — all on your machine. No cloud round-trip. No embedding API call. No telemetry on the code path. The engine is commercial software, but it runs entirely locally: nothing about your code goes anywhere we can see.
If you turn ArgosBrain off, your agent falls back to whatever it did before. No lock-in. The solo tier is free; teams pay for support and priority roadmap weight, not for the engine.
Get your free key → · The two services in detail · The two default skills · vs Cursor memories · vs Claude Code memory · Independent verdict