Skip to main content
By default the sandboxed process has unrestricted network access. When you restrict networking, nono routes traffic through an HTTP proxy running in the unsandboxed supervisor process. The sandboxed child can only reach localhost:<proxy-port> — all other outbound TCP is blocked at the kernel level. The same proxy can also be run on its own with nono proxy — a foreground server with no sandboxed child — so you can point external workloads (containers, microVMs) at nono’s domain filtering and credential injection.

Common cases

Proxy Modes

The proxy supports three modes that can be combined in a single session: When domain filtering or credential injection is active, the proxy starts automatically. HTTP_PROXY and HTTPS_PROXY are set in the child environment. For Node.js 26+, nono also sets NODE_USE_ENV_PROXY=1 so built-in fetch() reads HTTPS_PROXY. A 256-bit session token is generated for each proxy session. The child must include it in every request (Proxy-Authorization header for CONNECT, X-Nono-Token for reverse proxy), preventing other localhost processes from using the proxy.

Blocking All Network

No outbound connections are permitted. See Localhost IPC for allowing specific ports as exceptions. In a profile, set network.block to true:

CONNECT Tunnel (Domain Filtering)

The CONNECT tunnel is the most common proxy mode. It validates outbound hostnames against an allowlist, then relays raw TCP bytes. TLS is end-to-end between the sandboxed process and the upstream server — the proxy never sees plaintext. --allow-domain is an HTTP(S) proxy allowlist, not a generic raw TCP host:port allow. When domain filtering is active, direct outbound connections are blocked and traffic must go through the proxy.
When --allow-domain is given a URL with a path (e.g., https://github.com/org/**), only requests matching that path pattern are allowed for that domain. Requests to other paths on the same domain receive 403 Forbidden. A plain hostname (e.g., api.openai.com) allows all paths on the domain. Use nono why --self --host https://github.com/some/path inside the sandbox to check whether a specific URL is permitted and see the configured endpoint rules. Proxy architecture flow

DNS Rebinding Protection

The proxy resolves DNS itself and checks all resolved IPs against the deny list before connecting. This prevents DNS rebinding attacks where a malicious DNS server maps an allowed hostname to an internal IP.

Proxy Port

By default the proxy binds to an OS-assigned ephemeral port. Use --proxy-port to fix it when the application requires a known port:
Applications that read OPENAI_BASE_URL, ANTHROPIC_BASE_URL, etc. from the environment don’t need --proxy-port — nono sets these automatically.

Network Profiles

Network profiles are named sets of allowed domains composed from groups. They are defined in network-policy.json (embedded in the binary) and activated with --network-profile or via a profile’s network.network_profile field.

Preset Network Profiles

Groups

Each group maps to a set of allowed hostnames and wildcard suffixes:

Adding Domains

Use --allow-domain on the command line or allow_domain in a profile to add domains on top of the network profile:

Client-Side Proxy Bypass

Profiles can add host patterns to the generated NO_PROXY and no_proxy environment variables with network.no_proxy. nono still emits loopback defaults and any safe direct-connect bypasses first; parent-shell NO_PROXY values are not inherited. These entries do not grant network access; on macOS ProxyOnly and other restricted modes, direct connections still require matching sandbox permissions.
Entries must be host patterns, not URLs or host:port values. Bare multi-label domains and protected metadata suffix tokens are rejected because common clients treat them as suffix matches; use a safe single-label local alias like redis, a canonical IP literal, or an explicit suffix pattern such as .internal.example. Patterns conflicting with proxy-allowed hosts or credential/L7-filtered route upstreams fail configuration so that traffic stays on the local nono proxy. Generated NO_PROXY / no_proxy values use common environment-client syntax, so profile *.internal.example is emitted as .internal.example. nono also emits NONO_NO_PROXY with canonical nono patterns, preserving forms such as *.internal.example, so wrappers can translate them to runtime-specific formats like Java http.nonProxyHosts.

Reverse Proxy (Credential Injection)

When --credential is active, the proxy runs a reverse proxy alongside the CONNECT tunnel. The sandboxed process sends plain HTTP requests to http://localhost:<port>/<service>/..., and the proxy injects the real API key and forwards to the upstream over TLS.
The agent never sees the real credential. A phantom session token is placed in the sandbox environment; the proxy swaps it for the real key at the HTTP header level. See Credential Injection for full details on credential sources, custom credentials, and endpoint filtering.

External Proxy (Enterprise Passthrough)

For corporate environments with a mandatory outbound proxy, chain domain-filtered traffic through it with --upstream-proxy. This is typically paired with the enterprise network profile:
CONNECT requests are chained through the corporate proxy. Cloud metadata endpoints are still denied regardless.

Bypassing the External Proxy

Some domains may need to bypass the upstream proxy and connect directly:
Bypass patterns support exact hostnames and *. wildcard suffixes (case-insensitive). Matching hosts route directly; everything else goes through the upstream proxy. This can also be configured in a profile:
Or via environment variables:

Standalone Proxy (nono proxy)

nono run / shell / wrap start the proxy as a side effect of launching a sandboxed child, wiring HTTP_PROXY/HTTPS_PROXY and the session token into that child. nono proxy instead runs the same network-filtering / credential-injection proxy as a foreground server with no sandboxed child, prints the connection details, and blocks until you press Ctrl-C. This lets you point your own tools at nono’s domain filtering and credential injection, useful for workloads that provide their own OS-level isolation, such as Docker containers, Kata Containers, Kubernetes pods, Firecracker, and gVisor. The workload routes its egress through the proxy via HTTP(S)_PROXY while isolation comes from the container/microVM runtime.
On startup it prints the proxy URL (with the embedded credential), the raw token, ready-to-paste HTTP(S)_PROXY exports, a summary of configured routes, and — when TLS interception is active — the trust-bundle path:
Settings are loaded from --profile and extended/overridden by explicit flags, reusing the same configuration machinery as the sandboxed path. Allow-domains and credentials are merged on top of the profile; the network profile and upstream proxy are overridden by flags. Credential injection for known services and cmd:// credential-capture routes carry through the profile exactly as they do under nono run.

Flags

Reusing a CA across runs

By default each nono proxy run mints a fresh ephemeral CA, so any client that trusts the intercepted TLS has to re-trust a new certificate every time. To keep one CA across runs — trust it once in your clients, containers, or microVM images — supply your own with --proxy-ca-cert and --proxy-ca-key:
The key must be an ECDSA P-256 private key in PKCS#8 PEM (-----BEGIN PRIVATE KEY-----); the certificate must be a CA certificate in PEM. Generate a pair with OpenSSL:
The key and cert are validated at startup: a mismatched pair (or an unreadable file) fails the command immediately rather than silently disabling interception. Validity comes from the supplied certificate, so --proxy-ca-cert cannot be combined with --proxy-ca-validity; on macOS it also cannot be combined with --trust-proxy-ca, which manages its own CA in the Keychain. Protect the key file — anyone who can read it can mint certificates your clients will trust.
TLS interception (and therefore the CA) only activates when a route requires layer-7 visibility — for example an endpoint-scoped --allow-domain (a URL with a path glob) or --allow-endpoint rules. Plain host allowlisting tunnels CONNECT traffic without interception, so no CA is used.

Authentication

By default the proxy generates a random per-session token; clients authenticate with Proxy-Authorization (Basic via the nono:<token> userinfo in the printed URL, or Bearer). To use a fixed credential, pass --pass (or set NONO_PROXY_PASS to keep it out of shell history and process listings). --no-auth starts an open proxy that accepts every request on the bind address. To fail secure, it is refused for non-loopback bind addresses so you can’t accidentally expose an open proxy to other hosts, and it is mutually exclusive with --pass.

Endpoint Filtering

When credential injection routes traffic through the reverse proxy, you can further restrict which HTTP method+path combinations are allowed on a per-service basis. This enforces least-privilege at the API level — the agent can reach an allowed domain but only use specific endpoints.

CLI Usage

Use --allow-endpoint to restrict a credential service to specific patterns:
When any endpoint rules are configured for a service, requests that don’t match receive 403 Forbidden and are logged in the audit trail.

Profile Configuration

Endpoint rules can also be defined on custom credentials within profiles:
See Credential Injection — Custom Credential Definitions for the full custom_credentials schema.

Pattern Syntax

Path patterns use standard glob syntax (same as .gitignore and nono profile include patterns):

Localhost IPC

Use --open-port to allow bidirectional localhost TCP on a specific port (connect + listen). This enables IPC between sandboxed processes — for example, an MCP server in one sandbox and an AI agent in another.
--open-port works alongside domain filtering. Outbound to allowed hosts goes through the proxy; IPC stays on localhost:
This generates Seatbelt rules (macOS) or Landlock rules (Linux) that allow both connect() and bind() on the specified port, in addition to the proxy port. See CLI Reference for full details and platform limitations.

Listen-Only Ports

Use --listen-port when a sandboxed process needs to accept inbound connections but does not need to initiate outbound connections on that port (e.g., a server or gateway):
See CLI Reference for platform limitations.

Always-Denied Destinations

The following destinations are always blocked by the proxy, regardless of configuration. These cannot be overridden. Private network addresses (RFC1918: 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16) are allowed to support enterprise environments.

Platform Behavior

Linux

Network filtering uses Landlock V4+ per-port TCP rules. The sandbox restricts connect() to only the proxy port. All other outbound TCP is blocked at the kernel level. Requirements: Landlock ABI v4+ (Linux 6.7+)

macOS

Network filtering uses Seatbelt rules. The sandbox allows only (remote tcp "localhost:PORT") and denies all other network operations.

WSL2

The WSL2 kernel (6.6) ships with Landlock V3, which does not include TCP network filtering (V4 requires kernel 6.7+). Additionally, the seccomp-based proxy fallback is unavailable due to WSL2’s own seccomp notify listener. What works: --block-net (blocks all networking via SECCOMP_RET_ERRNO). What doesn’t work: Per-port filtering (--open-port, --listen-port) and proxy-based domain filtering (--allow-domain, --network-profile, --credential). Domain filtering is blocked by default on WSL2 — set wsl2_proxy_policy: "insecure_proxy" in your profile’s security config to opt in to degraded execution. See Credential Proxy on WSL2 for details. When Microsoft upgrades the WSL2 kernel to 6.7+, per-port filtering will activate automatically. See WSL2 Support for full details.

Audit Logging

All proxy decisions are logged via tracing:
Enable verbose logging to see proxy decisions:

HTTP/2 Support (--allow-http2)

By default, the proxy uses HTTP/1.1 with keep-alive connection pooling for all upstream connections. The --allow-http2 flag enables HTTP/2 multiplexing via ALPN negotiation:
It can also be set in a profile with network.allow_http2, which both nono run and nono proxy honor. When enabled:
  • Reverse proxy: Negotiates h2 with upstreams that support it, allowing multiple requests to be multiplexed over a single TCP connection. This eliminates per-request TLS handshake overhead (~50-100ms) for workloads with many concurrent requests to the same host (e.g., Maven/Gradle artifact downloads).
  • CONNECT intercept: Advertises both h2 and http/1.1 in ALPN, enabling gRPC and other HTTP/2 clients to negotiate h2 transparently through the proxy. Per-stream endpoint rules and credential injection are enforced at the HTTP/2 frame level.

Why Disabled by Default

HTTP/2 is opt-in because:
  • Some upstream servers or corporate proxies may not handle h2 correctly
  • HTTP/1.1 keep-alive is sufficient for most workloads
  • h2 multiplexing changes request ordering semantics, which may affect rate-limited APIs

Connection Pooling

Regardless of --allow-http2, the reverse proxy reuses TCP+TLS connections via keep-alive. With HTTP/1.1, connections are reused but requests are serialized per connection. With HTTP/2, multiple concurrent requests share a single connection via stream multiplexing. The pool provides DNS rebinding protection: resolved IPs are pinned after the initial filter check, preventing TOCTOU attacks on reused connections.

Limitations

  • No per-port filtering on macOS — Seatbelt cannot filter outbound by destination port
  • Domain filtering requires supervised execution — The proxy runs in the unsandboxed parent process, so nono wrap (Direct mode) is incompatible. Use nono run instead.
  • WSL2: no per-port filtering or domain filtering by default — See WSL2 section above

Next Steps