Skip to main content
Caveman has two configuration layers: feature configuration used by caveman wrap and agent shortcuts, and proxy configuration used by caveman start. Credentials belong in environment variables or provider-native credential stores. Do not put API keys in either configuration file.

Feature configuration

Global feature configuration lives at:
A project can add a restricted overlay at:
Inspect the resolved path and values with:

Keys and defaults

Project overlays may set think.toon, think.shrink, remember.*, and execute.*. They cannot change think.mode, think.core, or pixel settings. This prevents a checked-in project file from silently enabling a more invasive transformation mode.

Environment overrides

Environment variables take precedence over stored feature configuration. Use environment overrides for temporary sessions. Use caveman tools config set for durable operator choices.

Proxy configuration

Default proxy configuration path:
Set CAVEMAN_CONFIG to load another file.

Main fields

Accepted internal proxy modes are record, recommend, shadow, canary, active, compress, and pixel. Unknown values resolve to record. Operator-facing local workflows normally use record, compress, or pixel. CAVEMAN_MODE can override proxy YAML mode for caveman start.

Proxy environment

These are read by the proxy process itself, not by feature configuration. With CAVEMAN_AUTH_TOKEN set, every request must present the token in x-cave-api-key or Authorization: Bearer. The proxy consumes that header before resolving the provider credential, so the token is never forwarded upstream. /health/live, /health/ready and /metrics stay unauthenticated. See Deploy the proxy for a team.

Provider overrides

Provider entries can change public endpoint or regional information without putting secrets in YAML.
wire_dialect declares the wire grammar the mount’s upstream actually speaks. Empty (the default) keeps the shared OpenAI shape everywhere. Set anthropic only when the upstream answers the Anthropic Messages protocol on the mount. It changes two behaviors, both scoped to Anthropic Messages-path requests:
  • Usage accounting parses with the Anthropic dialect. Such an endpoint reports input_tokens exclusive of cache reads/writes, which the OpenAI-shape contradiction check misreads as malformed usage on every cache-warm response, so token accounting (and with it cache status and compression eligibility) silently drops those rows.
  • Compression zones follow the Anthropic Messages grammar, which keys the live/frozen boundary on the request’s own cache_control breakpoints. The OpenAI grammar would mark only the latest messages live, leaving the uncached post-breakpoint tail frozen or rewriting a block the provider already cached.
Routing, headers, and pricing keep the mount’s OpenAI-compatible identity in every dialect, and OpenAI-protocol paths on the same mount keep the OpenAI grammar. A named compat mount can forward additional provider headers explicitly:
Only that mount forwards the listed headers. Standard provider credentials, request framing, cookies, and Caveman control headers cannot be configured this way. Headers named by an inbound Connection field stay on that connection. The proxy publishes header names, never values, with its endpoint identity. OpenClaw and Pi keep a provider direct when required configured headers are absent from that contract. SDK session-affinity headers (session_id, x-session-id, x-client-request-id, x-session-affinity) are forwarded as standard end-to-end fields. Self-hosted private or loopback upstreams require an explicit CAVE_SSRF_ALLOWLIST entry. See Security and privacy.

Corporate networks: proxies and TLS inspection

Provider traffic honours the standard HTTPS_PROXY, HTTP_PROXY, and NO_PROXY variables by default, the same way curl, Python, and Node do. On a host that only reaches the internet through a corporate proxy nothing extra is needed. upstream_proxy changes that:
A URL pins one proxy (http://, https://, or socks5://) for provider requests without exporting process-wide proxy variables that the wrapped agent’s shell commands would inherit. In both modes localhost, loopback addresses, and NO_PROXY matches are dialed directly, so an allowlisted local model server keeps working next to a corporate proxy. CAVE_UPSTREAM_PROXY overrides the YAML value. When a proxy is selected for a destination, Caveman does not resolve that hostname itself; the proxy does, which is what makes this work on a network with no outbound DNS at all. HTTPS providers tunnel through the proxy with CONNECT, so a plain forward proxy never sees request bodies or credentials. A proxy that performs TLS inspection (below) terminates TLS itself and does see them. A proxy that accepts the connection and then answers nothing is bounded by CAVE_GATEWAY_RESPONSE_HEADER_TIMEOUT_MS (default 900,000 ms; 0 disables), which covers the wait for response headers only and never truncates a running stream. The proxy address itself needs no CAVE_SSRF_ALLOWLIST entry; see Security and privacy for what the guard still checks when a proxy is in use. TLS inspection (Zscaler, Netskope, and similar) presents provider certificates signed by a company root. Trust it by pointing ca_bundle or CAVE_CA_BUNDLE at the PEM file:
Bundles already exported for other toolchains are picked up too, so an environment set up for curl, Python, or Claude Code works unchanged: Every bundle is additive on top of the system store, so public providers keep verifying when a bundle holds only the private root. A ca_bundle that is missing, corrupt, or truncated fails startup rather than being half-trusted. An inherited variable that cannot be loaded for any of those reasons is skipped with a startup warning instead. The wrapped agent talks to the local proxy on loopback. If the agent itself reads HTTPS_PROXY (Claude Code does), that hop would otherwise be handed to the corporate proxy and time out. caveman run and caveman native enable add the gateway host to the agent’s NO_PROXY automatically, appending to whatever you already set and keeping your spelling of the variable. Set it yourself only when you launch the agent without caveman.

Provider credentials

The proxy preserves an inbound request credential. When an integration does not send one, supported providers can use their standard environment variables. Common examples include:
Amazon Bedrock supports its native authentication paths, including AWS credentials and supported bearer-token configuration. Prefer provider-native credential discovery over copying secrets into shell history. Bedrock resolves credentials in the AWS default chain order: environment keys (AWS_BEARER_TOKEN_BEDROCK, then AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / optional AWS_SESSION_TOKEN), then web identity (AWS_WEB_IDENTITY_TOKEN_FILE + AWS_ROLE_ARN, which is how IRSA works), then container credentials (ECS task role, EKS Pod Identity), then the EC2 IMDSv2 instance profile. On ECS, EKS, or EC2 you therefore grant the task, pod, or instance role bedrock:InvokeModel* and set no AWS keys at all.

Precedence summary

Feature configuration resolves from defaults, global file, allowed project overlay, then environment override. Proxy mode resolves from default, YAML, then CAVEMAN_MODE. Command flags can select an explicit session mode such as caveman wrap --off or --pixel. When resolution fails or a mode is unknown, request transformation fails safe: the runtime uses record or original-byte behavior instead of guessing.