2026-05-14 OpenClaw outbound HTTP/HTTPS proxy on a rented ProxyMac Mac mini: why SSH exports lie and how launchd should inherit them

If you operate OpenClaw on a rented Apple Silicon Mac mini in Hong Kong, Japan, Korea, Singapore, or the United States, your laptop may already exit the internet through a corporate HTTP(S) proxy—while the mini itself does not magically inherit that path. Operators routinely export HTTP_PROXY=... inside SSH, watch model calls succeed, then ship a launchd LaunchAgent that inherits none of those variables and blame “OpenClaw latency.” This article gives (1) a crisp audience filter, (2) symptom patterns that mimic AI outages, (3) a four-mode proxy matrix, (4) a nine-step launchd encoding runbook, (5) a scenario table for PAC, split tunnel, and TLS inspection, plus (6) NO_PROXY footguns for stdio MCP. Pair it with MCP environment vs launchd, headless SSH checklist, split tunnel VPN routing, and gateway token JSON. For break-glass UI prompts use VNC; for SKU math use pricing; for path diagnostics use help.

Who must wire outbound HTTP(S) proxy explicitly for OpenClaw

Start with teams where the mini’s default route reaches the public internet without the same inspection chain as developer laptops. That includes regulated banks, OEM vendors with global ZTNA clients, and hybrid shops that only proxy 443 to approved SaaS domains. If your security baseline mandates explicit proxy credentials for non-browser daemons, assume OpenClaw’s gateway process needs the same contract. Conversely, if the mini already receives a transparent forward proxy via router policy, you may only need NO_PROXY tuning—but you still prove it with packet captures or curl -v evidence, not folklore.

• Quantitative gate: if curl -I https://api.openai.com from SSH returns 407 without --proxy but succeeds with it, you are in the explicit-proxy bucket.
• Latency gate: add 180–320 ms to model round-trips when traffic hairpins through a central proxy on another continent—budget MCP timeouts accordingly.
• Parallel gate: corporate proxies often cap concurrent CONNECT tunnels at 30–80; size OpenClaw parallelism below that ceiling before declaring instability.

Symptoms that masquerade as model outages or “bad regions”

HTTP-layer failures rarely surface as a polite Proxy Authentication Required banner inside JSONL. You more often see TLS handshake timeouts, abrupt connection resets after exactly 60 or 120 seconds, or partial streaming where the first tokens arrive then the socket dies when the proxy rekeys. Another mimic is DNS success with TCP hang—common when split DNS sends api.* to internal resolvers that only exist on the VPN interface your mini forgot to bring up. Before you move workloads from JP to SG, prove whether the failure reproduces on loopback curl with identical proxy flags.

Decision matrix: static proxy vs PAC vs transparent vs none

Mode	How OpenClaw sees it	launchd implication	Primary risk
Static http://host:port	Libraries read HTTP_PROXY/HTTPS_PROXY	Must copy into plist EnvironmentVariables	Secrets in world-readable plist if chmod wrong
PAC / WPAD URL	Runtime may ignore unless app resolves PAC	Often need wrapper exporting resolved proxy	PAC JS changes without notice—pin versions in Git
Transparent intercept	No env vars; MITM presents corporate CA	Trust store must include CA in system keychain	Custom CA rotation breaks TLS until renewed
None (direct)	Public egress from mini region	Still set NO_PROXY for localhost MCP	Accidental exfil if future policy adds proxy

Nine-step runbook: from SSH proof to launchd parity

1. Inventory variables: collect HTTP_PROXY, HTTPS_PROXY, ALL_PROXY, NO_PROXY, and any vendor-specific keys your IT names.
2. Prove curl parity: run curl -v --proxy http://proxy:8080 https://api.anthropic.com on the mini and archive the verbose log.
3. Encode plist env: mirror the winning values into EnvironmentVariables; avoid referencing shell-only ~/.zshrc exports per MCP env guide.
4. Split secrets: keep embedded basic-auth out of tickets—use launchd-friendly secret files referenced from plist as in token JSON article.
5. NO_PROXY hygiene: include 127.0.0.1,localhost,*.local plus RFC1918 ranges your MCP uses.
6. VPN alignment: if laptops rely on full tunnel, confirm mini routes with split vs full tunnel guide before chasing provider SLAs.
7. launchctl smoke: launchctl kickstart the agent and diff JSONL outbound lines against SSH foreground.
8. Soak test: run 500 sequential model calls at 3 concurrent workers to surface proxy connection caps.
9. Rollback: keep previous plist revision in Git; bootout and restore within 2 minutes if error rates spike.

Corporate egress scenario table (pick your row before tuning timeouts)

Scenario	Observable signal	OpenClaw tuning lever
Central EU proxy, mini in US	RTT adds > 220 ms vs direct	Raise HTTP client timeouts; lower parallel in-flight
Zscaler with SSL inspection	TLS alert unknown CA	Install corporate root in System keychain + reboot trust cache
Split tunnel omitting mini subnet	DNS resolves, SYN never completes	Fix VPN client route table; mirror laptop policy on mini
Proxy auth rotation every 12h	407 bursts at shift change	Use managed credential helper or short-lived tokens

Pitfalls: NO_PROXY mistakes that break stdio MCP while models look “fine”

stdio MCP servers bind to 127.0.0.1 high ports. If NO_PROXY is empty, some HTTP stacks still try to tunnel loopback through the proxy—resulting in bizarre half-working toolchains where the model answers but every filesystem tool times out. Conversely, an overly broad NO_PROXY=* disables the proxy entirely and can violate compliance. Prefer explicit comma lists and test with curl --noproxy '*' http://127.0.0.1:8765/health versus proxied calls. Document the final string beside your plist in the same repo that tracks headless checklist outputs so auditors see intent.

export HTTP_PROXY=http://proxy.corp.local:8080 HTTPS_PROXY=http://proxy.corp.local:8080 NO_PROXY=127.0.0.1,localhost,169.254.169.254

FAQ

Does OpenClaw honor lowercase vs uppercase env vars? Many libraries accept both; pick one convention per fleet and stick to it—mixed case doubles confusion during incidents.

Should SOCKS proxies use ALL_PROXY? Often yes for tools that are not HTTP-aware; verify with the same curl matrix you used for HTTPS.

What if IT forbids storing proxy passwords on disk? Use short-lived PAC or negotiate header injection via sidecar helper; still avoid interactive prompts on headless fleets—see VNC only for one-time trust prompts.

Why ProxyMac Mac mini is the right place to harden OpenClaw egress

Rented Mac mini M4 nodes give you native macOS trust stores, predictable single-tenant CPU for TLS-heavy proxy hops, and region choice across HK / JP / KR / SG / US so inference control planes sit closer to provider endpoints than a lone EU bastion. You keep SSH for receipts, help articles for cross-region playbooks, and pricing transparent when you need another mini to isolate prod vs staging egress policies—without buying five closets of hardware.