Caveman
Caveman is a middleware layer that reduces AI agent operating costs by compressing tool-result tokens—such as logs, JSON, and diffs—before they reach the model, and by adaptively selecting the cheapest model that maintains output quality. It works with 30+ coding agents like Claude Code and Codex, as well as custom TypeScript and Python agents, and offers a cloud gateway with telemetry and automated improvement loops. The platform claims to shrink an 18,240-token test log to 201 tokens while preserving original content for on-demand recovery.
- Developer Tools
- Artificial Intelligence
- AI Agents
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
AI agent runs incur significant costs from large input token volumes, particularly from verbose tool results like logs, test output, JSON, and diffs. Teams also struggle to select the optimal model for each task, often overpaying for frontier models on routine work, which inflates operational expenses without proportional quality gains.
Solution
Caveman reduces the cost of every agent turn by shrinking input tokens and adaptively selecting the cheapest model that gets each task right. It operates as a middleware layer that compresses tool results before they reach the model, while keeping originals recoverable on demand. For off-the-shelf agents like Claude Code, Codex, and OpenCode, Caveman cuts log and test output dramatically—for example, shrinking an 18,240-token test log to 201 tokens. For custom TypeScript or Python agents, it integrates into existing frameworks without changing providers, keys, or saved history. Caveman Cloud adds a managed gateway with telemetry, evaluation, and improvement automation that observes live traffic, tests candidate fixes, and proves savings with labeled evidence (measured, inferred, tested, or verified).
Target Audience
Primary customers are engineering teams running AI agents at scale—both those using off-the-shelf coding agents like Claude Code and Codex, and those building custom TypeScript or Python agent applications—who need to control LLM costs without sacrificing output quality.
Features
- Token compression for logs, test output, JSON, and diffs, with original content recoverable on demand in supported agent loops
- Adaptive model selection that routes each turn to the cheapest model and effort level that maintains correctness
- Works with 30+ coding agents including Claude Code, Codex, OpenCode, and Gemini CLI, plus custom TypeScript and Python agents
- Local runtime that sits beside the app without changing the framework, provider, or saved history
- Caveman Cloud gateway with provider-compatible endpoints, request-level tracing, and workload-level observability
- Automated improvement loop that observes traffic, diagnoses waste, generates candidate changes, evaluates them against test cases, and deploys only after passing eval gates
- Evidence labeling system that distinguishes measured spend, inferred savings, and verified savings, with verified savings starting at $0 until proven
- Read-only ClickHouse SQL access for querying requests, spans, tool events, evals, and logs