Max out your models.Buff your work.

Models keep leapfrogging — keep your go-to agent.Pool every AI subscription & API to run Claude Code, Codex, and Antigravity at full capacity.

Download ModelBuffRead the docsmacOS · Windows · Linux
ClaudeChatGPTGeminiKimiGLMDeepSeekClaude CodeCodexAntigravity
The ModelBuff 6-section desktop workbench: Live topology, Agents, Models & pools, Media MCP, Network exits, and Usage analytics.

Live routing topology, agents, dual-window pools, chained exits, and usage in one workbench — click to explore.

Built for yourentire AI stack.

Plans & APIs
  • Claude
  • ChatGPT
  • Gemini
  • Antigravity
  • Cursor
  • Kimi
  • GLM
  • DeepSeek
  • MiniMax
  • Grok
  • Volcengine Ark
Image & video
  • RunningHub
  • LibTV
  • Dola
  • Muse
  • Higgsfield
  • Kling
  • Seedance
  • Seedream
  • Veo
  • GPT Image
Agents & clients
  • Claude Code
  • Codex
  • Antigravity
  • Cursor
  • OpenCode
  • Windsurf
  • Gemini CLI
  • Cherry Studio
  • Claude Desktop

01Smart scheduling

Dual-window quota,smart load balance.

Smarter than round-robin: every request picks the best account, and rate limits hand off live with zero interruption.

  • Expiring weekly firstSpend accounts closest to their 7-day reset first so quota never expires unused.
  • Fresh windows heldKeep unused 5-hour accounts on standby until needed so timers do not start early.
  • Load balance & relaySpread active load by session count and usage; skip full accounts and hand off live on rate limits.
running non-stop_Account Ain usecooling downstandbyAccount Bin usecooling downstandbyAccount Cin usecooling downstandby

02Precise routing

Directory, group, agent:multi-axis routing.

Swap in the newest model without switching agents, while directory bindings keep work and personal accounts apart.

  • Directory bindingEntering a bound folder locks in its account automatically — run your CLI as usual.
  • Group boundariesDefine isolated account groups so scheduling and failovers stay strictly inside.
  • Keep your agentSwitch to newer models or ordered fallback chains across providers without changing tools.
Claude CodeWork projectWork projectPersonal projectPersonal projectWork accountPersonal account

03Image & video

Web media credits,as standard MCP tools.

Break image and video credits on RunningHub, LibTV, Dola, and Muse out of isolated web canvases into any MCP agent.

  • Web credits to MCPExpose web-only image and video credits locally as dedicated standard MCP servers.
  • Inline generationCall image and video generation directly from Claude Code, Codex, and Antigravity — no web clicking.
  • Saved to your projectGenerated assets land straight in your workspace instead of staying locked inside a web canvas.
RunningHubLibTVDolaMuseClaude CodeCodexAntigravityMCP

04Own space

One closed loop each.No overlap, no fallback.

Fixed exit, clean sandbox, risk check, account browser & official CLI — stops if a boundary breaks.

  • Exit & clean sandboxPer-hop load balancing · Pinned exit per account · Fail-closed
  • Environment checkExit & IP reputation · DNS & leaks · Region & sandbox
  • Browser & native CLINative browser · 0700 profile · Official CLI
Clean sandbox · 070001Fixed exityour nodes · fail-closed02Clean sandbox0700 dir · TZ & locale03Health checkexit IP · DNS · latency04Isolated loginnative browser · 070005Official CLIby folder · native CLI

05Official CLI

Run the official CLI.No proxies or cloaking.

Third-party proxies and cloaking invite account bans — ModelBuff drives the official CLI natively inside a clean sandbox.

  • Always currentNothing bundled or version-pinned — ModelBuff runs the official CLI installed on your machine.
  • Open official protocolsIntegrate through standard ACP and MCP instead of brittle, unofficial proxy layers.
  • Zero prompt tamperingRequests go straight from the official CLI with no cloaking or prompt rewrites, avoiding proxy ban risks.
ACPMCPofficial build

06Featherweight

Lightweight sandbox.Just one computer.

Lighter than VMs or Docker: no extra OS, no images, no cloud hosts.

  • Under 5 MB RAMOne resident process under 5 MB per session.
  • ~45 ms launchFull sandbox + fixed exit adds ~45 ms (~3 ms identity switch).
  • No remote hostsReplace VMs, Docker, and cloud servers with one laptop.
VM · DockerCloud serverYour computer~45 msVMDockerVMCloud serverCloud server

07Redaction gateway

Local redaction,one unified gateway.

Listens on 127.0.0.1:28330 across major API formats — redacts secrets outbound and restores live on return.

  • Multi-protocol & keysSpeaks OpenAI, Anthropic, Gemini & more, with scoped local keys and quotas.
  • Secret maskingAPI keys and private keys are masked before leaving your machine.
  • Reversible aliasesPaths, emails & regex rules swap outbound and restore live without breaking cache.
Your computerModel providerOpenAIAnthropicGemini127.0.0.1:28330127.0.0.1:28330sk-abc123…user@example.comuser_1@example.invalid~/work/app/workspaceredacteduser@example.comuser_1@example.invalid~/work/app/workspaceredactedrestored

08Quota radar

Live quota radar,every account in view.

Track dual-window balances, reset countdowns, and consumption pace across every subscription pool on one screen.

  • 5h & 7d live countdownsRemaining quota and exact reset timers for both rolling windows stay updated in real time.
  • Five live pool statesPriority, active, standby, cooling down, and weekly cap update continuously across your pool.
  • 7-day pace benchmarkCompare real usage against an even-pace reference line to spot fast burn or idle expiring quota.
Account poolClaude · account 1in use72%Claude · account 2standby100%ChatGPT · account 1resets in 2h18%Gemini · account 1in use46%

09Usage analytics

Full usage analytics,every token accounted.

Combine routed gateway traffic and native local CLI sessions in one private on-device ledger.

  • Gateway + native sessionsAggregates local gateway calls alongside direct Claude Code, Codex, Antigravity, and OpenCode sessions on device.
  • Cache savings & offsetBreak down input, output, and prompt cache hits to quantify subscription offset alongside API spend.
  • Speed & 52-week heatmapCompare first-token response and output speed alongside a 52-week activity heatmap.
This weekBy modelBy agentBy dayhighesthighesthighestClaudeClaude CodeGPTCodexGeminiAntigravityMTWTFSS

10Trace & context

Route & context X-ray,turn by turn.

Make agent calls transparent: inspect every routing handoff, latency segment, and context window breakdown.

  • Routing topologyInspect candidate scoring, skipped full accounts, live rate-limit failovers, and the active exit node.
  • End-to-end latency splitPinpoint bottlenecks across redaction, connection setup, first-token wait, and streaming output.
  • Context & cache diagnosisSplit system prompt, tool definitions, tool results, and history shares plus cache hits and turns left.
Turn 1read 3 filesTurn 2ran 2 toolsTurn 3edited 4 filesTurn 4all tests passSystem prompt11%Tools14%Tool results45%Conversation30%cache hit92%0%22%41%63%84%Context

Give your agentsa full Buff.

curl -fsSL https://go.modelbuff.com/install.sh | sh

macOS · Windows · Linux100% local runtime. No cloud account required. MIT-licensed open source.