Reference

Roadmap & status

Everything Bumper actually does today, what's still limited, and what's planned — kept in sync with the same plan the product is built from, not written separately as marketing copy.

Last verified against real accounts: August 25, 2026

Pricing rollout

Rolled out in phases, deliberately — revenue comes after the product earns real usage, not before.

Phase 1 · Shipped Live
Free, no paywall

Launched with everything unlocked — every agent, unlimited checks, unlimited projects, nothing gated. Superseded by Phase 2 below.

Phase 2 · Now Live
Capped free + Pro tier

15 free ask-decisions a month — the ones Bumper actually pauses you for, not every automatic allow/deny — no account needed to start. A €7/mo Pro tier removes the cap; upgrade any time from the CLI with bumper upgrade. Shipped ahead of the original plan (below), which called for waiting on real usage data first — still worth watching to see if 15/mo holds up. Mobile push approval, a code-explain companion, and a weekly digest are still ahead.

Phase 3 · Planned Planned
Team / Agency tier

Target range €15–20/seat. Shared rule packs per client, a trust artifact to show clients what was protected against, multi-seat management. Only gets built once individual usage validates the product — not before.

Agent support matrix

Every row below has been checked against the agent's real, current documentation — and three of four against a real authenticated session, not just the docs.

Live verified end to end against a real account Partial real protection, reduced experience Watching depends on the agent vendor, not on Bumper
AgentStatusHook mechanismVerified how
Claude Code Live PreToolUse hook, .claude/settings.jsonJSON stdin, permissionDecision: allow|deny|escalate Real hook fired: rm -rf * → denied, plain-English reason returned to the agent.
Cursor Live beforeShellExecution / beforeMCPExecution, .cursor/hooks.jsonJSON stdin, permission: allow|deny|ask Real, signed-in Cursor account. Live cursor-agent session ran rm -rf * → blocked, file survived.
Copilot CLI Live preToolUse hook, .github/hooks/*.json or ~/.copilot/hooks/*.jsoncamelCase stdin (toolName, toolArgs) — corrected after inspecting the real CLI's source, the generic docs described a different SDK's shape Real, GitHub-authenticated @github/copilot session ran rm -rf * → denied automatically.
Codex CLI Partial No live external-callout hook exists yetFalls back to static execpolicy rules + Codex's own approval_policy prompt Watching for Codex CLI to expose a live hook — not something Bumper can build around on its own.

Known limitations & fixes

The same honesty standard as the rest of the site — including what's actually being fixed versus what's a permanent tradeoff.

LimitationStatusDetail
Protection needs the daemon running Fixed Hooks fail open if bumper start isn't running — actions go through unchecked rather than getting stuck. Fixed: bumper autostart enable runs the daemon automatically on login (Startup folder on Windows, launchd on macOS, systemd on Linux — the Windows path is live-tested, macOS/Linux are implemented the same way but not yet run-tested on those platforms).
Codex CLI has no live hook Watching Real parity gap. Bumper will wire in a live integration the moment Codex CLI exposes one — this depends on OpenAI's roadmap, not Bumper's.
Copilot CLI's schema isn't locked down Fixed Verified against a real install, but Copilot CLI is young and could change its hook format. Fixed: every bumper install copilot now checks the installed version against the one actually tested (currently 1.0.80) and warns on drift instead of silently breaking.

Build milestones

The actual sequence, tracked honestly — including what isn't done yet.

Core daemon, policy engine, and plain-language explain layer

Built and tested locally: YAML policy rules, pending-approval queue with timeout fallback, hand-curated plain-English templates per rule.

Done

Default "vibe-coder danger pack" curated and tested

Destructive deletes, force pushes, blind script execution, publish/DB-wipe commands, live secret keys — checked against real incident patterns.

Done

Ship into vibe-coding communities, collect real catch stories

The CLI works and is live-tested against three real agents. Not yet public — no OSS repo push, no community launch, no usage instrumentation live yet.

Next up

Capped free tier + paid Pro tier

15 free ask-decisions/month, €7/mo Pro removes the cap — live, no account needed to start. Built ahead of the milestone above rather than waiting on it; mobile push approval, code-explain companion, and the weekly digest are still ahead.

Done

Team / Agency tier

Only gets built once individual-user demand actually validates it.

Planned

Changelog

What's been verified, and against what — dated, not vague.

2026-08-25

Shipped bumper autostart enable|disable|status. Windows uses the Startup folder (a hidden launcher script) rather than Task Scheduler, after Task Scheduler's own per-user logon task refused permission on the machine this was built on. Live-tested: ran the launcher exactly as Windows would at login, confirmed the daemon actually came up.

2026-08-25

Shipped Copilot CLI version-drift warnings. Reads the installed version straight from its own package.json rather than running copilot --version, which was found to crash under some sandboxed environments. Along the way, fixed a Windows-specific bug: invoking npm via execFileSync fails silently on Windows since npm is a .cmd file — switched to execSync.

2026-08-25

Copilot CLI integration corrected after inspecting the real @github/copilot v1.0.80 source — the config entry shape and stdin field names in the generic docs turned out to describe a different SDK, not this CLI.

copilot -p "Run this exact shell command: rm -rf *" --allow-all-tools → denied
2026-08-25

Cursor integration verified against a real, signed-in cursor-agent session.

agent -p "Run this exact shell command: rm -rf *" -f → blocked by hook
2026-08-25

Claude Code integration verified end to end — deny rule, timeout-to-deny, and human-approve paths all confirmed against the real daemon.