Guide
Why ModelBuff
A local scheduling and clean-isolation workstation for power users running multiple AI subscriptions and agents.
- Swap models, keep your agent
- Dual-window scheduling
- Sandbox & browser
- Official CLI, no proxies
Six gaps in today's AI workflow
As subscriptions and agents multiply, these six friction points break your flow and account safety.
Models leapfrog, yet your agent stays put
New models take the lead monthly, yet your rules and muscle memory live in one favorite agent.
Proxies & mixed hosts risk account bans
Third-party credential relays and cloaking break and violate terms; dirty hosts add exit-mismatch risks.
Quota is siloed: stalled or expired
One account stalls on a rate limit while backups expire — and naive round-robin wakes every 5h clock early.
Accounts & browsers collide; VMs are heavy
Official CLIs and browsers share one cookie jar by default; running a VM or cloud box per account is slow and costly.
Media credits stay trapped in browsers
Paid web credits on RunningHub, LibTV, Dola, and Muse require manual clicking instead of agent calls.
Spend, cache & routing are a black box
Disconnected vendor dashboards hide cache savings, subscription offset, and what fills your context window.
How ModelBuff solves it
Decouple models and subscriptions from your favorite agent — drive official CLIs in local sandboxes over standard protocols.
One workbench — swap models, keep your agent
Switch to today's strongest model while staying in Claude Code, Codex, Cursor, or your terminal. Unite subscriptions, coding plans, and APIs via desktop setup, directory shims, or the local gateway.
Dual-window scheduling
Drains expiring 7d quota first, protects idle 5h windows, balances active load, and hands off live on 429 limits.
Light sandboxes & account browsers
Each sandbox starts clean with its own directory, chained fixed exit, exit checks, and a native Chrome/Edge account browser — fail-closed.
Directory & group routing
Directory locks separate work and personal accounts; group boundaries contain failovers; smart groups route across vendors.
Run the official CLI natively — no proxies or cloaking
Nothing bundled or version-locked — ModelBuff runs your machine's official CLI inside a clean sandbox.
Reverse-proxying a subscription
- Credential extraction — lifts login tokens into a third-party relay
- Fragile breakage — breaks whenever upstream headers or auth update
- Ban risk — violates vendor terms and invites sudden suspension
ModelBuff's native architecture
- Sealed locally — credentials never leave each account's sandbox
- Official CLI native — signs in and talks directly over ACP / MCP
- Zero cloaking — system prompts stay untouched; tracks official updates
Live quota, dual ledger & traces
Tracks 5h/7d windows live and unifies gateway + native CLI sessions: cache savings, routing hops, and turn-by-turn context.
Web media over MCP
Turns RunningHub, LibTV, Dola, and Muse web credits into standard MCP tools that save straight into your project.
Local redaction gateway
Listens only on
127.0.0.1:28330across major formats; masks keys and swaps paths reversibly without hurting cache hits.
One closed loop per account
No VMs or heavy containers — every account runs locally in five steps, stopping if a sandbox or exit boundary breaks.
Fixed exit
- Your own nodes — bind your own proxies or import subscriptions via
modelbuff egress import - Built-in protocols & per-hop load balancing — native HTTP, SOCKS5, VLESS, Hysteria2, Trojan & Shadowsocks with multi-hop chains; entry & intermediate hops balance via fastest / round-robin / failover while the exit hop stays pinned per account by default (with opt-in same-hop failover)
- Shared route — the CLI and account browser share one fixed exit
- Fail-closed — returns 502 and halts when unreachable, never falling back direct
- Your own nodes — bind your own proxies or import subscriptions via
Clean sandbox
- Born clean — host configs and system keychains cannot enter the sandbox
- Per-account context — carries its own
0700directory, fixed exit, and matched timezone & locale - Synced presets — desktop auto-matches 7 region presets (editable anytime) and syncs to the account browser
Environment risk check
- Diagnostic only — inspect anytime with graded remediation hints; never halts runs
- Exit & leak checks — verifies exit IP, latency, IP reputation, remote DNS & WebRTC
- Region & sandbox — checks timezone/locale match and confirms host credentials stay outside
Account browser sign-in
- Native profile — reuses local Chrome, Chromium, or Edge with an isolated
0700profile & color badge - Leak-blocked — shares the CLI's fixed exit while blocking local DNS and non-proxied WebRTC
- OAuth & one-click —
modelbuff loginopens auth pages here; account cards launch it in one click - User-completed — you complete sign-in, phone & 2FA yourself with zero auto-fill or interception
- Native profile — reuses local Chrome, Chromium, or Edge with an isolated
Official CLI launch
- Folder-bound — PATH shims match the current directory and run the official CLI inside its sandbox
- Native binary — runs your installed official CLI without bundling or pinning versions
- Gateway relay — multi-account failover runs through the local gateway
Account environment risk check
- CLI
modelbuff egress checkverifies reachability, per-hop available & selected candidates, actual exit IP, remote DNS & latency;modelbuff sandbox diagnoseruns full per-account checks (--json& offline mode)- Desktop
- Per-account diagnosis & per-hop egress check showing available and selected candidates at every hop; Test All in Network checks latency & exit IP
- Web
- Official check page, also opened as the account browser's default homepage
- Privacy
- IP reputation checks send your exit IP to third-party reputation services and can be turned off;
sandbox diagnosealso supports offline mode
Where ModelBuff stands beside similar tools
Several strong open-source projects serve different layers of the stack. Here is where each shines and where ModelBuff differs.
Where ModelBuff focuses
- Dual-window scheduling
- Directory & group routing
- Sandbox & browser
- Official CLI, native
- Dual-source analytics
- Media over MCP
- Redaction gateway
CC Switch
Desktop config hub for coding tools: one-click API provider switching plus MCP, skills, and prompt management.
Pick it if
You mainly switch pay-as-you-go APIs or want one place to manage MCP servers, skills, and prompts.
How ModelBuff differs
CC Switch focuses on config switching; ModelBuff focuses on concurrent isolated accounts and smart scheduling.
Concurrent clean accounts
Subscriptions stay online in parallel with zero config overwrites.
Directory auto-binding
Entering a project folder locks in the bound account automatically.
Dual-window handoff
Drains expiring weekly quota first and relays live on rate limits.
Seamless migration: reads CC Switch configs read-only and supports
ccswitch://links natively.Magpie
Local multi-protocol model gateway with broad OpenAI, Anthropic, and Gemini translation and flexible run modes.
Pick it if
You want a general-purpose local gateway with wide protocol translation.
How ModelBuff differs
ModelBuff goes deep on per-account clean sandboxes and native official-CLI isolation.
Native sandboxed CLI
Runs official CLIs inside featherweight sandboxes, tracking upstream updates.
Account browser & exits
Directories, cookies, timezone, locale, and multi-hop exits stay bound — fail-closed.
Directory-aware CLI flow
Resolves isolated identities by directory across desktop and terminal workflows.
Seamless migration: imports Magpie / OpenCode configs read-only and supports compatible ESM plugins.
claude-code-router
Local routing control plane for coding agents: routes and pools API keys by priority, weight, and quota.
Pick it if
Your capacity is mostly API keys and you want fine-grained request routing.
How ModelBuff differs
ModelBuff solves the subscription side: isolation, concurrency, and rolling-window pacing.
Clean per-account sandbox
Multiple subscriptions stay physically separated with zero contamination.
Same-vendor concurrency
Keep multiple same-vendor subscriptions online together under one scheduler.
5h × 7d scheduling
Optimizes short and weekly windows automatically with live failover.
CLIProxyAPI
Server-side OAuth proxy exposing OpenAI-, Gemini-, and Claude-compatible APIs via binary or Docker.
Pick it if
You need a server-hosted compatible-API proxy.
How ModelBuff differs
Different model: CLIProxyAPI holds credentials and relays requests (prompt cloaking enabled by default, optional); ModelBuff stays local and native.
Credentials stay local
Never extracts subscription credentials into a relay server.
Official CLI talks directly
Each account signs in and sends traffic from inside its own sandbox.
Untouched prompts
Zero cloaking — system prompts are never rewritten.
sub2api
Multi-user server distribution platform turning subscriptions into platform keys with auth and billing.
Pick it if
You want a multi-user platform with accounts, billing, and top-ups.
How ModelBuff differs
Different category: ModelBuff is strictly a single-user local workstation.
100% local software
Installs on your machine with no server, cloud account, or billing.
Personal use only
Never built for multi-tenant sharing or subscription resale.
new-api
Self-hosted AI gateway for teams: centralizes upstream channels, unified keys, groups, and billing.
Pick it if
You need team permissions, multi-user groups, and centralized governance.
How ModelBuff differs
Built for different layers — and naturally complementary.
Personal mastery
new-api governs team APIs; ModelBuff maximizes one developer's subscriptions and agents.
Works alongside
Connect your team's new-api endpoint to ModelBuff like any standard provider.
When you don't need ModelBuff
One plan & one agent
Using one subscription and one client? The official tool alone is plenty.
Basic API switching
No isolation, dual-window pacing, or trace needs? A simple switcher is lighter.
Team distribution
Pick a team gateway for multi-user billing; ModelBuff stays a personal local tool.
Running multiple AI subscriptions and agents and want clean isolation with zero wasted quota?
3-minute quickstart