Skip to main content
Caveman Cloud makes every AI agent think, remember, and execute on a fraction of the cost, and proves it. These questions come up most often from the three teams that evaluate, adopt, and govern the product. Each answer links to the relevant documentation for deeper reading.
Where the honest answer is “not yet” or “it depends,” we say so. We do not claim controls or outcomes we do not have.

​
For developers

No. Compression through a proxy is net-negative on latency; we never promise faster. What we promise is cheaper, with proof. Caveman routes your request to the provider and returns the response unchanged. When an optimization like compression runs, the transform happens inline; when the cache store is unreachable, the request forwards upstream normally, costing you the miss but never the call.See How It Works for the full request lifecycle.
Local tools (caveman start, caveman wrap) need no account and keep working if Cloud is unreachable. For hosted gateway traffic, if the provider is unreachable, the gateway returns its own error with a Caveman request ID so you can trace it. Cache outages cost you the hit, never the call. If an optimization fails during processing, the gateway falls back to the original request and continues.See Troubleshooting for exit codes and diagnosis steps.
The gateway changes the bytes the model sees only when the request asks for it. Prompt-cache hints run on every request and never change model-visible text. Compression runs only when a request asks for it with x-cave-optimize: compress, and then rewrites tool output the model sees, losslessly or with an encrypted original for recovery. record mode is always pure pass-through. Unknown modes fail closed to the most conservative handling.See Rollout Safety for how changes are gated before they reach production.
By default we keep the redacted prompt and response text until you delete it or set a window, envelope-encrypted in private object storage, so your traffic can be replayed to prove a fix before it ships. You can switch Raw payload storage off in the console to store metadata only: spans, token counts, cost, latency, and salted hashes. On Free, payload storage cannot be switched off. Send x-cave-retention: metadata or x-cave-retention: zdr on any request to override per-call.See Data and Privacy for the full consent and retention model.
It depends on your plan. On Free, training use is a disclosed condition of the plan and may not be switched off. On Pay-as-you-go, training use is opt-in and off by default. On Enterprise and customer installs, training use cannot be turned on; the control API refuses. No plan’s data is used for training in production today.See Data and Privacy for the plan-conditioned policy.
Caveman works with OpenAI, Anthropic, Google, Azure, and OpenAI-compatible endpoints. You can use the OpenAI SDK, Anthropic SDK, Vercel AI SDK, LangChain, or LiteLLM. For coding agents, Claude Code, Codex, Gemini CLI, and OpenCode are supported.See the integrations for OpenAI, Anthropic, Google and HTTP, Frameworks, and LiteLLM.
Caveman is a base-URL swap. Change your client back to the provider’s native base URL and remove the x-cave-upstream-key header if you were using per-request keys. If you stored provider keys in Caveman, rotate them at your provider after removal. No code changes are required beyond the configuration revert.See Quickstart for the original one-line integration pattern.
Measured spend is observed usage priced at public catalog list prices. Inferred is an estimate from the engine’s local counterfactual; it is not a guaranteed invoice reduction. Verified is reserved for per-request causal proof where a Caveman transform caused a provider-measured delta, backed by one of three supported methods.See Savings Evidence for the full contract.
Verified savings can be zero for honest reasons: the request used record mode (pass-through), the provider or model is not on the allow-list for verified methods, a blocker such as output-shape prevented count-baseline verification, the model is unpriced in the catalog, the call failed, or it was a cache hit with no counterfactual provider response to compare.See Troubleshooting for the complete list.
Use request headers: x-cave-optimize: off for pass-through, x-cave-optimize: compress to ask for compression, x-cave-cache: semantic,ttl=900 for semantic cache with a 15-minute lifetime. Unrecognized values return 400 with a clear error code.See Control Optimizations for the full header catalog.
Your Cave API key authenticates you to Caveman Cloud and determines your tenant scope. For metadata reads (traces, reports), you need trace:read_metadata. For billing columns, you need billing:read. For payload reads, you need payload:read plus trace:read_payload.See Authentication for key types and scopes.
Yes. Run read-only ClickHouse SELECT over your project’s requests, spans, tool events, and evaluations. Agent connections and sql:read keys are limited to 101 rows, 128 KiB, and 15 seconds. Human queries get up to 10,000 rows, 8 MiB, and 30 seconds.See Query with SQL for schema details and examples.

​
For engineering leaders

Caveman is a one-line integration: swap the base URL and keep your provider key. It works with agents your team already runs (Claude Code, Codex, Gemini CLI, OpenCode) and agents you ship (OpenAI SDK, Anthropic SDK, Vercel AI SDK, LangChain). If you already run LiteLLM, its OpenTelemetry exporter sends traces to Caveman directly or through your OTel Collector; provider credentials stay in LiteLLM.See Connect a Workload and Connect a Coding Agent.
Caveman pays off when input is much larger than output and the input is repetitive or bloated: RAG with fat retrieved chunks, agents re-sending big system prompts and tool schemas, tool and log-heavy pipelines, and high-volume support bots with long preambles. It is a weaker fit for short prompts with long outputs, already-lean prompts, and quality-critical work that cannot be evaluated.The benchmark envelope is roughly 20 to 30 percent potential input-cost reduction on input-heavy workloads. This is not a verified claim or invoice guarantee.See Optimization Catalog for what each optimizer does and how it is gated.
Caveman provides soft and hard caps per project, with an Inbox item and email when a soft cap is crossed. Rate limits apply per key and per workflow. Positive hard budgets block traffic at admission: the gateway reserves a conservative maximum before the upstream call and refuses when settled spend plus live reservations would exceed the cap. Monthly soft spend alerts remain advisory and fail open.See Budgets and Guardrails for setup and policy configuration.
Five roles (owner, admin, engineer, viewer, billing) with explicit permission tables. Sensitive scopes are narrowed: raw payload read is owner/admin only, publishing aggressive policies is owner/admin only, and connecting a repo is owner/admin only. An admin cannot mint or modify owner access. SSO via OIDC and SAML is supported. An append-only audit log records administrative and security-relevant actions inside the same transaction as the change.See Team and Admin for role details and SSO setup.
Yes. For enterprise deployments, Caveman runs in your VPC on AWS or GCP via Terraform and Helm, using your KMS and storage, with content sharing locked off. A local, single-tenant option (caveman start / wrap) also runs on your own machine with no account required, though it stays inference-only with no verified-savings claims.See Deployment Options for hosted, your cloud, and local deployment paths.
Every query carries an explicit organization filter from verified auth, never from the request body. Postgres has row-level security enabled and forced on all tenant tables. Storage uses organization-scoped prefixes, tenant-bound AEAD encryption, and composite foreign keys. Cross-tenant access is denied and tested at every layer.See Security Review for the full isolation architecture.
Every candidate change runs through an evaluation suite before it can be proposed. Evaluations check quality, cost, and behavior against test cases drawn from your traffic. A change that does not pass stays out of production. Approved changes roll out through the gateway with traffic-share controls and safety criteria you configure.See Rollout Safety and Evals for the gating process.
The local proxy (caveman start / wrap) sends usage telemetry to Caveman by default, but you can opt out with caveman telemetry off, CAVEMAN_TELEMETRY=0, or DO_NOT_TRACK=1. The CLI telemetry includes an install ID and client IP, but never prompts, responses, code, or file paths. Customer-owned cloud deployments keep operational telemetry in your own ClickHouse instance.See Deployment Options for air-gapped and customer-cloud details.
Start with the local proxy on your own machine to see inferred estimates on your actual traffic without routing production through a managed gateway. When you are ready, connect a non-production workload to the hosted gateway and run evaluations against your test cases. Every improvement is proposal-only until your team approves it.See Adoption Playbook for a phased evaluation plan.
You control deletion. Set retention windows to auto-delete old data, switch Raw payload storage off to purge all captured bodies, or delete the organization to remove everything. Backups roll off over about 30 to 32 days after the purge, so total removal from every copy takes up to about 62 days.See Data and Privacy for retention and deletion controls.

​
For finance and procurement

Caveman is a platform fee plus usage-based compute. We do not publish prices in documentation. Contact contact@caveman.so to discuss pricing for your workload, or book a conversation at cal.com/caveman/chat.See Reporting Savings for how savings are tracked and reported internally.
Caveman keeps three numbers separate. Measured spend is observed usage at public catalog list prices; it is not your provider invoice. Inferred is an estimate with a daily range and confidence, never re-projected to a month. Verified is a per-request causal delta backed by provider-reported usage, counted only on eligible providers and models where the causal contract is proven.Verified savings can be negative on write-heavy days; the ledger stores signed values without flooring to zero.See Savings Evidence for the full accounting contract.
Verified savings start at the honest zero. They require a supported accounting method (provider causal cache, provider causal cache on Bedrock, or provider-counted baseline delta), a priced model in the catalog, provider-complete usage, and a qualifying request. Many workloads spend days or weeks accumulating before the first verified figure appears. This is expected and honest, not a bug.See Troubleshooting for the detailed breakdown.
Use the Evidence reports in the console, which label every figure with its basis: measured, inferred, or verified. Verified savings are the only per-request causal number and the only one suitable for audit. Inferred headroom is useful for internal prioritization but should not be reported as realized savings. Measured spend gives you a catalog-price subtotal for cost tracking.See Reporting Savings for report types and export options.
Caveman replaces no existing vendor. It sits between your applications and your model providers, adding telemetry, governance, and optimization layers. You keep your provider relationships and keys. The platform fee is predictable, and compute usage scales with your traffic. No gainshare or savings-linked billing mechanics exist today.See How It Works for where Caveman sits in your stack.
Hosted Caveman Cloud runs on GCP in europe-west4 with ClickHouse Cloud in the same region. Subprocessors include Google Cloud, ClickHouse Cloud, Cloudflare, Resend, PostHog, Stripe, Supabase, Vercel, Google Workspace, Cal.com, and GitHub. Your upstream model providers reached via BYOK are your own processors, not Caveman subprocessors.See Security Review for the complete subprocessor table and region details.
Enterprise deployments run under a separate agreement that defines region, access, backups, and deletion responsibilities. For custom terms, contact contact@caveman.so.See Deployment Options for enterprise and customer-cloud details.
This FAQ and the Security Review page are designed to answer the questions that show up on vendor security questionnaires. For questions not covered here, or for the latest legal documents, email contact@caveman.so.
Yes. View the interactive demo with fictional data at caveman.so/demo.

​
Still have questions?

Security Review

Deep architectural and compliance review for InfoSec and procurement teams.

Troubleshooting

Fix common issues with authentication, traces, gateway URLs, and savings labels.

Contact Us

Email contact@caveman.so for questions not covered here.

Book a Call

Schedule a conversation to discuss your workload and requirements.