Caveman Cloud· HostedGet Started

Caveman Cloud: see what your agents cost and cut it

Caveman Cloud records the agents your team runs and ships, ties their cost to the work they did, and counts a saving only once production proves it.

Caveman Cloud shows what your agents cost and cuts it, with proof. It records the coding agents your team runs and the agents you ship, ties what they cost to the work they did, finds waste, and tests each proposed change against your own traffic before it ships. Traffic reaches it three ways: point an SDK at the Caveman gateway, route your coding agents through it, or send OpenTelemetry traces from a gateway you already run, such as LiteLLM.

Caveman Cloud is self-driving software for your LLM stack. Built-in agents observe your traffic, evaluate candidate changes, prove them with evidence, and propose improvements. You stay in control and approve what ships. The loop runs continuously: Observe, Diagnose, Generate, Evaluate, Deploy, Learn.

What sets Caveman apart is that it proves its own claims. Every number carries a label (measured, inferred, tested, or verified), changes ship only after they clear eval gates, the default path is byte-safe, and verified savings start at an honest $0. Read Why Caveman for the full picture.

Choose your path#

Developers

Integrate in one line, label traffic, control optimizations per request, and debug with traces, SQL, and evals.

CTOs and Platform Leads

Govern AI spend across teams, agents, and coding tools with eval-gated rollouts, RBAC, and audit.

CFOs and Finance

Understand what each number means, which savings are verified, and how to report them with confidence.

30-Day Adoption Playbook

Connect, diagnose, prove, and report in four weeks.

Security Review

Hosting, encryption, key custody, tenant isolation, and data policy.

Deployment Options

Hosted Cloud, your own VPC, on-prem, or local-only tools.

Let AI drive#

Connect Your Coding Agent

Give Claude Code, Codex, or Cursor access to your project over MCP.

Automated AI Setup

Paste one prompt and let your coding agent wire up the gateway, evals, and baseline.

Built-in AI

See how Ask Caveman, Improvements, judges, and rollouts work together.

Improvements

Review agent-generated changes backed by Evidence reports before they ship.

Get started#

Quickstart

Send your first request through the Gateway and see it in Traces within minutes.

Connect a Workload

Route agent or application traffic with a base-URL swap and two headers.

How It Works

Understand the Observe to Learn loop and how Caveman Cloud handles your traffic.

OpenAI Integration

Point the OpenAI SDK (TypeScript or Python) at the Caveman gateway.

TypeScript SDK

Trace workflows, search tools, and export telemetry from your application.

CLI

Query traces, run SQL, and operate the control plane from the terminal.

Onboarding#

Sign in to the console
Open your Caveman Cloud console and sign in. If you were invited to a project, you land directly in that workspace.
Create a project and API key
Go to Gateway, Connect to create a project, then generate a Cave API key. Copy the key and your gateway URL. The gateway URL has no trailing slash.
Set environment variables
Export CAVE_API_KEY with your Cave key and CAVE_GATEWAY_URL with your gateway origin. Keep your upstream provider key separate.
Swap the base URL
Change the baseURL on your OpenAI or Anthropic client to ${CAVE_GATEWAY_URL}/openai/v1 or the matching provider path. Add the x-cave-upstream-key header when the provider key is not stored in Caveman Cloud.
Send a request and view the trace
Run one chat completion or agent turn. Open Traces in the console to inspect the request, model, usage, and latency.

Core surfaces#

Surface	What you do with it
Gateway	Route traffic, apply optimizations per request, and record usage with provider-compatible endpoints.
Traces	Inspect every request, its cost, latency, and outcome. Search by agent, workflow, model, and time.
Workloads	Register and observe the agents and applications sending traffic to your project.
Improvements	Run improvement attempts against recorded traffic, evaluate changes with test cases, and review Evidence reports.
Ask Caveman	Investigate traces and workloads with an agent-assisted conversational interface.
Automation	Configure monitors and rules that watch traffic and surface decisions in the Inbox.
Inbox	Review what Caveman noticed, what it did, and what needs your decision across every subsystem.
SQL	Run read-only ClickHouse SELECT over your project's requests, spans, tool events, evals, and logs.

Savings evidence#

Caveman Cloud reports three kinds of numbers, and it keeps them separate:

Measured spend records observed usage and its pricing basis.
Inferred savings describe an estimate or counterfactual, not a guaranteed invoice reduction.
Verified savings require a supported accounting method and qualifying evidence. They stay zero when that evidence is absent.

See Savings Evidence for the full contract, Optimization Catalog for what each optimization changes, and Reporting Savings for turning these numbers into a monthly finance report.