Skip to main content
VibeDefend is the automated path to Cybe MCP — plus everything around it. The Cybe MCP page walks you through adding the MCP URL to each client by hand. VibeDefend does that for you, and goes further: it wires a hook layer so the right business and security rules arrive in the agent’s context at the moment it edits code, hard-blocks denied actions, and runs a gap analysis at the end of each substantive session. One command, every detected agent.

​
Install

npx -y @cybedefend/vibedefend@latest install
Works on macOS, Linux and Windows — PowerShell, cmd, bash, zsh, fish, Git Bash, pick any. Requires Node 18.17 or later, which most users already have because Claude Code / Cursor / Codex ship a bundled Node. The installer is fully interactive: pick a region, pick which agents to wire (they’re auto-detected), confirm. That’s it. Prefer a global install? 
npm install -g @cybedefend/vibedefend && vibedefend install
# pnpm and yarn work the same way

​
MCP alone vs. MCP + hooks

Connecting the MCP server by hand (the Cybe MCP flow) gives your agent a passive security backend: it can list findings, drill into a vulnerability, drive the status lifecycle and pull business-logic context — when you ask it to. VibeDefend adds the active layer on top. The hooks fire on the agent’s own lifecycle events, so security shows up without anyone prompting for it:
MomentWhat VibeDefend doesWhy it matters
Before an edit (Edit / Write / MultiEdit)Injects the business + security rules relevant to this change into the agent’s context.The agent writes secure, on-policy code from the first line instead of being corrected afterwards.
On any tool callAction Guards evaluate the call against your guard policy and hard-block on deny.A rm -rf, a write to a protected path, or a fetch to a blocked host is stopped before it runs — not flagged after.
At session startLoads the CybeDefend doctrine and opens the proposals inbox (Accept / Reject picker).The agent starts every session knowing your rules, and you clear pending rule proposals up front.
At session end (Stop)Runs a gap analysis once the session crosses the edit threshold.Surfaces the business rules you relied on but never wrote down, and proposes them for review.
On long sessions (PreCompact)Same gap analysis, fired before the context window is compacted.Long sessions don’t lose their rule-gap review just because the transcript got summarised.
Per prompt (Claude Code)A lightweight doctrine backstop on every user prompt.Keeps the CybeDefend flow intact even when another skill auto-activates and would otherwise take over the turn.
The MCP server is what the hooks call under the hood. VibeDefend installs both, so you get the tools and the moments that use them.

​
Supported agents

VibeDefend auto-detects and wires whichever of these you have installed. Unchecked agents stay untouched — re-run vibedefend install any time to toggle one on or off.
CapabilityClaude CodeCursorOpenAI CodexWindsurfVS Code Copilot
MCP server install✅✅✅✅✅
Business + Security Rules (injected pre-edit)✅✅✅⚠️ writes only✅
Action Guards (hard block on deny)✅ all tools✅ all tools✅ all tools⚠️ writes + MCP fallback¹❌ not yet wired
Session Start (loads doctrine + proposals inbox)✅✅✅⚠️ proxied²✅
Session Review (end-of-session gap analysis)✅✅✅✅✅
PreCompact (long-session gap analysis)✅✅❌ no event❌ no event✅
Doctrine backstop (per-prompt reminder)✅❌✱ via MCP³❌❌
Min versionlatest≥ 1.7latestlatest≥ 1.110
Legend — ✅ supported · ⚠️ supported with caveats · ❌ not exposed by the agent · ✱ alternate mechanism ¹ Windsurf’s pre_write_code hook hard-blocks on file writes only. For non-write tool calls (Read / Bash / WebFetch) the installer drops a snippet into .windsurfrules instructing the agent to call cybe_guards_check via MCP before sensitive actions — soft enforcement that relies on the model following its rules file. ² Windsurf has no native SessionStart event. VibeDefend wires pre_user_prompt, which fires on every turn. The hook is idempotent and cheap (one GET to the proposals endpoint, returns “0 pending” once the inbox is empty), so the per-turn cost is negligible. ³ Codex follows the doctrine via the MCP server’s Server.instructions field on each session, which removes the need for a per-prompt reminder hook.

​
The install flow, step by step

vibedefend install walks you through four prompts and then signs you in.
1

Pick a region

Choose Production — EU or Production — US. VibeDefend registers an MCP server named cybedefend pointing at that region’s endpoint and wires the hooks to call the matching Gateway — you never type a URL. The two regions are independent CybeDefend tenants; sign in to the one that hosts your account (eu.cybedefend.com vs us.cybedefend.com).
2

Pick which agents to wire

VibeDefend prints a detection table (one row per agent, with a version badge) and a multi-select. Detected, supported agents are pre-checked; a version that’s too old is shown disabled with an “update this client” hint; an agent you don’t have yet is still selectable, so you can pre-wire hooks before you install it.
3

Tune the hook config

Four questions, each with a tuned default:
  • Enable the session-end gap analysis? — default on. Without it, the gap analysis never runs.
  • Edits threshold to trigger it? — default 3. Typos and one-line fixes won’t fire it; a real feature will. Lower to 1 to fire on any session that had an edit, raise it if your team does frequent tiny commits in long sessions.
  • Enable automatic background updates? — default on. Hooks self-refresh so you never re-run install (see Keeping VibeDefend current).
  • Enable auto-propose mode? — default off (not recommended). Off means rule proposals need an explicit chat-side “yes”. On means the agent drafts and pushes proposals without asking first — you still review them at the next session start.
4

Sign in

If VibeDefend has no stored credentials, it runs the login flow automatically: OAuth 2.0 Authorization Code + PKCE opens your browser to the CybeDefend sign-in page, receives the callback locally, and stores the token bundle in your OS keychain. If your browser already has a dashboard session, this completes instantly. The refresh token persists for 14 days.
After install, drop one tiny file at the root of each repo you want monitored — a .cybedefend/config.json with your project UUID: 
{ "projectId": "<your-cybedefend-project-uuid>" }
Grab the UUID from the project page on the EU dashboard or the US dashboard. You can also override it per shell or per project with the CYBEDEFEND_PROJECT_ID environment variable.

​
Commands

vibedefend install            Set up MCP + hooks (interactive)
vibedefend update             Upgrade the binary if newer, then refresh hooks
vibedefend update --self      Upgrade only the CLI binary (don't touch hooks)
vibedefend status             Read-only install report (region, agents, API check)
vibedefend doctor             Diagnose the install and repair what's fixable
vibedefend login              (Re-)authenticate against the CybeDefend API
vibedefend --help             Full help
vibedefend doctor takes --check for a dry-run (report only, change nothing) and --yes to apply every fix without the confirmation prompt. vibedefend login takes --force to re-authenticate even when credentials are already stored.

​
Keeping VibeDefend current

At install you’re asked “Enable automatic background updates? (recommended)”. Leave it on and you never update by hand: the SessionStart hook checks npm at most once a day and, when a newer version is published, upgrades @cybedefend/vibedefend in the background — the next session picks it up via the self-resolving runner shim. Opt out (or pin a version, or run in a locked-down environment) and you instead get a one-line nudge. To update manually at any time: 
vibedefend update     # upgrades the binary if newer, then refreshes hooks
(vibedefend update --self upgrades only the binary, without touching hooks.)
One-time bootstrap for installs older than 1.2.0. Those versions shipped a frozen copy of the hook runner with no update logic, so vibedefend update alone can’t upgrade the binary. Do it explicitly once:
npm i -g @cybedefend/vibedefend@latest   # or: vibedefend update --self
vibedefend update                         # render the self-resolving shim
After that, updates are hands-off (or one vibedefend update) as above.

​
Tunables

The gap-analysis threshold and auto-propose mode are chosen at install and stored in your VibeDefend config. To change them, re-run vibedefend install and answer the hook-config prompts again:
SettingWhat it controls
Review thresholdEdits before the session-end gap analysis fires (default 3; 1 fires on any session that had an edit).
Auto-proposeWhether the agent proposes rules without asking first (default off).
One override is read live from the environment, so you can set it per shell or per project — handy when a repo’s UUID differs from the default:
VariableEffect
CYBEDEFEND_PROJECT_IDOverride the project UUID from .cybedefend/config.json.

​
Codex setup gotcha

Codex 0.131+ requires you to approve each hook from the /hooks panel inside Codex before it fires. After running vibedefend install, open Codex, run /hooks, and trust the cybedefend entries — until you do, the panel shows Installed N / Active 0 and the hooks stay dormant.

​
VibeDefend or the manual MCP setup?

Both connect your agent to the same CybeDefend tenant. The difference is how much of the platform comes along.
VibeDefendManual MCP setup
SetupOne command, all detected agents at onceEdit each client’s config file by hand
MCP tools (read findings, triage, context)✅✅
Rules injected before each edit✅—
Action Guards (hard block)✅—
Session-end gap analysis✅—
Best forDay-to-day coding where you want security in the loop automaticallyAny MCP client VibeDefend doesn’t auto-wire (Continue, Cline, Zed, Claude Desktop…), or when you only want the read/triage tools
You can also do both: run VibeDefend for the agents it supports and add the raw MCP URL by hand to any other client.

​
Troubleshooting

Codex gates hooks behind manual approval — see Codex setup gotcha above. Open Codex, run /hooks, and trust the cybedefend entries.
Run vibedefend status for a read-only report of region, wired agents and a live API check, then vibedefend doctor to diagnose and repair what’s fixable (vibedefend doctor --check reports without changing anything). Confirm the repo has a .cybedefend/config.json with a valid projectId — without a linked project the rule-fetch and gap-analysis hooks have nothing to call.
Expected, once. Codex’s MCP-scoped token (aud: cybedefend-mcp) has limited Permify rights and can’t call the Gateway endpoints the runtime hooks use, which need a user token (aud: cybedefend-api). They aren’t interchangeable. One Logto sign-in (instant if your browser already has a dashboard session) issues the right token; the refresh token then persists for 14 days.
Decline the “automatic background updates” prompt at install (or it’s disabled automatically when VibeDefend can’t self-resolve a global install). You then get a one-line nudge when a new version ships and update on your own schedule with vibedefend update.
Fully supported — no Git Bash or WSL required. The hook runner is a single bundled Node script that runs the same way on Linux, macOS and Windows.

Cybe MCP, connect every AI agent

The manual setup and the full reference for the 18 MCP tools VibeDefend wires up for you.

Cybe MCP, feature overview

The product-level pitch and where the MCP layer sits in the platform.

Cybe Security Champion

The agent behind the business-logic rules VibeDefend injects before each edit.

Cybe Autofix

The fix-generation pipeline the agent can drive once it’s connected.