Caveman

Freemium

Caveman 是一套面向 AI 智能体和生产工作流的 token 效率工具栈，帮助团队精简符合条件的上下文、路由流量，并衡量经验证的节省效果，支持本地工具、TypeScript SDK 和托管云层。

什么是 Caveman？

Caveman is an efficiency stack for agent-native development and AI operations. It combines local compression tools, a TypeScript agent SDK, and an in-development cloud gateway so teams can reduce eligible context, track token spend, and verify where savings come from.

The product is organized as layered components that can be adopted separately. The site presents the Caveman Skill for output compression in Claude Code and 30+ agents, Caveman Proxy for recoverable local context compression, and Caveman Agent SDK for setting spend guards and context rules in TypeScript. Caveman Cloud and Caveman Enterprise extend the same system to hosted and on-prem traffic with eval-gated optimization and verified savings workflows.

Caveman 能做什么？

Output compression for agents

The Caveman Skill rewrites agent replies into shorter outputs while keeping code, commands, and errors byte-for-byte exact. The source describes a roughly 65% average reduction in output tokens for supported prompts.

Recoverable local context compression

Caveman Proxy runs locally and compresses eligible context on the user’s machine without requiring an account. The product text says the original bytes are recoverable.

TypeScript agent controls

Caveman Agent SDK lets developers define tools, context, evals, and spend guards in TypeScript so each run can follow declared checks before a cheaper context plan is locked in.

Spend visibility and catalog pricing

The workspace shows AI spend by cause, model, key, member, and workflow, using provider-reported usage against public catalog list prices rather than guessed averages.

Eval-gated optimization

Caveman Cloud applies caching, compression, and routing only when traffic qualifies. The site emphasizes fail-closed behavior and rollback when gates do not pass.

Verified receipt and ledger workflow

The pricing page describes a causal-cache ledger, receipt verification, receipt export, and Ed25519 verification in preview, with automatic signing and gainshare charging disabled.

使用场景

“Reduce verbose agent output”

Teams using Claude Code, Codex, Cursor, Gemini, or similar tools can install the skill to make responses shorter while preserving exact code and command content.

“Compress local agent traffic”

Developers who want lower token usage without sending prompts to a hosted service can use the local wrap/Proxy flow on their own machine.

“Control agent spend in production”

Teams building support, operations, or document agents can define context, tools, and evals in TypeScript and keep only context plans that pass their checks.

“Measure and explain AI costs”

Organizations that need a clearer spend view can use the workspace to break costs down by member, key, model, workflow, and cause.

“Pilot hosted optimization later”

Teams that start local can move toward Caveman Cloud or Enterprise when they want eval-gated caching, routing, verified savings, or on-prem deployment.

常见问题

What does Caveman actually do?

Caveman provides tools for compressing AI output, reducing eligible context, applying spend guards, and measuring AI usage across local and hosted workflows.

Is the whole stack available in one product?

The site presents Caveman as a stack of related products. Some parts are live, while Caveman Cloud and Enterprise are described as in development or preview.

Does the local proxy require an account?

No. The pricing page says the local wrap runs on your machine with BYOK and no account required, and a free account is optional for cloud sync and a dashboard seat.

Which agents are supported by the skill?

The source names Claude Code, Codex, Gemini, Cursor, and 30+ agents.

Can original context be recovered after compression?

The product pages describe recoverable compression for the proxy layer and byte-safe handling for supported transformations, with original bytes stored before lossy replacement.