Private CA and outbound proxy
Make a self-hosted Skilder trust your own certificate authority and send its outbound calls through a forward proxy.
Pilot release. Self-hosting Skilder is a pilot.
SKILDER_CA_BUNDLE makes Skilder trust the certificates your own certificate authority issued. HTTPS_PROXY, HTTP_PROXY and NO_PROXY send its outbound HTTP calls through your forward proxy. Each works alone, and a proxy that inspects TLS needs both.
Both apply to backend, remote-mcp and main-runtime, and to the processes main-runtime starts: MCP servers run with npx or uvx, and skill scripts.
Private certificate authority
A TLS certificate is issued by a certificate authority (CA). Skilder trusts the public CAs that browsers trust. When a service it calls has a certificate your company's own CA issued, the call fails as untrusted, with an error such as unable to verify the first certificate.
You need a CA bundle when Skilder calls one of these over TLS with such a certificate:
- an MCP server, or a package registry mirror that
npxoruvxinstalls from; - an OAuth authorization server on a public address (OAuth and DNS);
- your mail server;
- your forward proxy, when it inspects TLS, or when its URL starts with
https://and your CA issued its certificate.
A CA bundle is a text file in PEM format that holds one or more CA certificates. Skilder trusts them on top of the public CAs, which it keeps trusting.
Build the bundle
Ask your security team for the certificate of the CA that issues your internal certificates, and of any intermediate CA below it. To see which CA issued a service's certificate, run this on any machine that reaches it:
openssl s_client -connect internal.example.com:443 -servername internal.example.com -showcerts </dev/nullEach certificate of the chain prints an s: line, its subject, and an i: line, the CA that issued it. The bundle needs the CA on the last i: line, and each CA above it up to the root.
Each certificate in PEM format starts with -----BEGIN CERTIFICATE-----. A .cer or .der file in binary form converts with openssl x509 -inform der -in corp-ca.cer -out corp-ca.pem. Put the certificates one after the other in one file, then list what it holds:
cat corp-root-ca.pem corp-issuing-ca.pem > ca-bundle.pem
openssl crl2pkcs7 -nocrl -certfile ca-bundle.pem | openssl pkcs7 -print_certs -nooutIt prints a subject= line per certificate. bad end line means a file lacks its final line break, so Node would not load the bundle in full and the services refuse it at boot: add the line break (echo >> corp-root-ca.pem) and run both commands again.
Set it up
Copy the bundle next to .env
Keep ca-bundle.pem in the skilder directory, next to .env, and make it readable by everyone (chmod 644): the services do not run as root. Use a copy rather than a file of your host's own trust store. On a host with SELinux, Docker relabels the file for the containers.
Set the path
Set its absolute path, which realpath ca-bundle.pem prints. A relative one resolves against the directory of compose.yaml.
SKILDER_CA_BUNDLE=/srv/skilder/ca-bundle.pemSet SKILDER_CA_BUNDLE, not NODE_EXTRA_CA_CERTS: compose mounts the file in the three services and sets that one itself.
Recreate the services
Run docker compose up -d. After you change its contents, docker compose up -d alone leaves them running: use docker compose up -d --force-recreate --no-deps backend remote-mcp main-runtime.
A path with no file fails docker compose up. A file the services cannot read, that holds no certificate or that holds a broken one stops each of the three at boot. docker compose logs backend names the problem, for example NODE_EXTRA_CA_CERTS (/etc/skilder/ca-bundle.pem) holds no PEM certificate.
npx, npm, uvx, uv, Python and curl in the processes main-runtime starts trust the bundle too. A browser that an MCP server drives, such as Playwright's Chromium, reads its own trust store and does not. An MCP server whose own configuration sets a CA variable, such as SSL_CERT_FILE or NODE_EXTRA_CA_CERTS, keeps its own value.
Check the bundle
Call one of your internal services from the backend container, with https://internal.example.com replaced by its address:
docker compose exec backend node --use-env-proxy -e "fetch('https://internal.example.com').then((r) => console.log(r.status), (e) => console.log(e.cause?.message ?? e.message))"It prints the HTTP status. --use-env-proxy sends this one-off call through your outbound proxy, when you set one. Replace backend with main-runtime or remote-mcp to check those.
unable to verify the first certificate, unable to get local issuer certificate, unable to get issuer certificate or self-signed certificate in certificate chain means the bundle lacks the CA that issued the service's certificate, or a CA above it.
Outbound proxy
When your network lets nothing out except through a forward proxy, set its URL in .env:
HTTPS_PROXY=http://proxy.example.com:3128
HTTP_PROXY=http://proxy.example.com:3128
NO_PROXY=.corp.example.com,mcp.internal.example.comRun docker compose up -d to apply a change. Compose takes a variable your shell exports over the one in .env: when your shell sets HTTPS_PROXY, HTTP_PROXY or NO_PROXY, give it the same value. npm, uv and curl in the processes main-runtime starts read the same variables.
HTTPS_PROXYcarries the calls tohttpsaddresses andHTTP_PROXYthe calls to plainhttpones. Set both, usually to the same URL.HTTP_PROXYalone carries both.- Write the scheme,
http://orhttps://: any other value stops the three services at boot. Withhttps://, the services check the proxy's own certificate: a CA that issued it must be public or in your bundle. - Credentials go in the URL, as
http://user:password@proxy.example.com:3128. Percent-encode a special character in them, such as@as%40. The MCP servers and skill scriptsmain-runtimestarts can read them, so give Skilder a proxy account that grants nothing else. NO_PROXYlists the hosts reached directly, such as your internal MCP servers, separated by commas: a name or an IP address, a domain suffix that starts with a dot, or ahost:port. It takes names, not address ranges.
For https calls, your proxy sees a CONNECT to the requested host and port, so a proxy that allows destinations by hostname works. HTTPS defaults to port 443, but an address a member or admin enters can name another, such as https://mcp.example.com:8443: allow that port too. An MCP server whose own configuration sets HTTPS_PROXY, HTTP_PROXY or NO_PROXY keeps its own values.
Traffic that stays off the proxy
compose.yaml puts the stack's own services at the start of NO_PROXY, and adds yours after them:
localhost,127.0.0.1,proxy,nats,dgraph-zero,dgraph-alpha,object-store,backend,remote-mcp,main-runtime,frontendThe services call each other by these names, which resolve on the Compose network only. Your proxy cannot resolve them, so a call sent to it would fail, and internal traffic would leave the host. Your NO_PROXY adds to this list and cannot remove from it.
These calls never use the proxy:
- the backend's mail, which is SMTP, not HTTP. Allow the SMTP host directly;
- the analytics the
frontendcontainer forwards to PostHog, when you setPOSTHOG_KEY. Allow its hosts directly; gitand Docker on your host, which do not read.env.gitreadsHTTPS_PROXYfrom your shell, so clone with the token over HTTPS: SSH ignores it. Docker pulls through the proxy in its daemon configuration.
Outbound access lists every host and how each is reached through a proxy.
OAuth and DNS
Before the backend calls an OAuth authorization server or an MCP client's client_id URL, it resolves and checks every address. It refuses private network addresses with or without a proxy, regardless of the certificate's CA. The backend still needs DNS access to resolve public names when you use a proxy.
Through the proxy, these calls then ask for the hostname like any other https call, and TLS verifies that name. Keep your proxy's default deny of private and internal destinations. OAuth servers must use https in production: Skilder sends a plain-http one to the proxy by its checked IP address, which a proxy that allows by hostname only refuses.
A proxy that inspects TLS
A proxy that decrypts https traffic presents certificates that your own CA issued. Besides the proxy variables, put the CA your proxy signs with in SKILDER_CA_BUNDLE. Without it, every https call through the proxy fails as untrusted, the npx and uvx installs included.
git and Docker on your host go through that proxy too: add the same CA to the host's own trust store, then restart Docker.
Check the proxy
Call a public address from the backend container:
docker compose exec backend node --use-env-proxy -e "fetch('https://registry.npmjs.org/').then((r) => console.log(r.status), (e) => console.log(e.cause?.message ?? e.message))"It prints 200 when the call went out through the proxy and the proxy allowed it. Your proxy's access log shows a CONNECT registry.npmjs.org:443. --use-env-proxy makes this one-off node process read the proxy variables, which Node ignores on its own. The services set this up themselves.

