Give Claude a real terminal — without giving Claude the keys.
Claude runs the safe tasks for you — you stay in control of the rest. Reading files, checking logs, running builds, deploying, diagnosing errors — Claude handles those directly, no copy-pasting required. High-risk operations stay permanently blocked and in your hands by design. Every call is checked against a three-tier security model, and every decision is logged.
Two plugins, one trust model
local-terminal
- 8 structured tools — file listing, read, search, git, npm, system info, escape hatch.
- .mcpb extension — installs directly into Claude Desktop. No Windows Service or configuration files required.
- 140+ RED patterns across 27 categories block deletion, shutdown, code exec, exfiltration, credential access, persistence.
- Sensitive files protected —
.env,.ssh, credential stores blocked even from read tools.
vps-control
- 17 structured tools — PM2 status, deploy pipeline, git, stdout & error logs, file reads, system health, audit log.
- Supabase bearer token auth with subscription plan validation, TLS via sslip.io + Let's Encrypt out of the box.
- 116+ RED patterns across 21 categories block deletion, reboot, shell invocation, DB writes, package install, privilege escalation.
- Symlink realpath check and path allowlist keep file reads inside declared directories.
How we keep Claude from doing the wrong thing
Every command Claude tries to run passes through a three-tier classifier before the plugin executes anything. Structured tools execute fixed allowlisted operations (no user-provided command string reaches the security classifier); only the run_command escape hatch accepts arbitrary command text. The pipeline runs in order: Layer 1 hard-block static patterns (HARD_BLOCKED_PATTERNS, 140+ patterns across 27 categories) reject instantly; then — if an Anthropic API key is configured — two parallel AI safety checks (Layer 2 + Layer 3); then the remaining RED static pattern checks (BLOCKED_PATTERNS); then the AMBER warning classifier. No single tool bypasses the check. Audit logs go to disk with token and secret redaction. Without an Anthropic API key (or if the API call fails), the AI classification layers are silently skipped and the plugin falls back to the static RED block list plus AMBER dry-run warnings only. Set LAYER_STRICT_MODE=true to block on API failure instead of skipping.
RED — hard-blocked
File deletion, shutdown, code exec via PowerShell/bash -c, curl/wget/scp/ssh, database writes, privilege escalation, Unicode homoglyphs, newline injection (each line is checked independently against the block list). No override, no flag, no escape for the hard-blocked categories.
AMBER — dry-run default
Bulk copies, find -exec, sed -i, wildcard renames. dry_run=true is the default; an AMBER match fires a visible warning. Passing dry_run=false on the first call executes immediately — the recommended workflow is to preview first, then confirm.
GREEN — allowed & logged
Everything else. Timeout-capped per call, redacted audit log, and a dry-run-first gate on the escape-hatch tool. The boring tier by design.
Pricing
All plans include a 14-day free trial — no charge during the trial period. Subscriptions are otherwise non-refundable except for confirmed ForgeRift billing errors, prorated convenience-termination refunds, and applicable statutory consumer rights — see Terms §6.5.
⚠️ local-terminal-mcp: Windows 10 / 11 only. vps-control-mcp is accessible from any OS that runs Claude Desktop. macOS support for local-terminal-mcp is planned for a future release.
- All tools for your chosen plugin
- Email support available
- All local-terminal tools
- All vps-control tools
- Email support available
- Individual plugin — $9.99/mo rate-locked while subscription stays active
- Bundle — $14.99/mo rate-locked while subscription stays active
- Monthly billing only — no annual option
- First 100 paid subscribers or 3 months after the marketplace listing date — whichever comes first