Skip to contentdocs.Growther.si
Download C5

Network and egress

C5 runs entirely on your computer, but it does reach out for a few things: your licence, updates, the small models some features run on your machine, and — if you use Remote Control with a tunnel — the tunnel program. This page is the complete account of what it contacts, when, how to route that through your proxy, and how to turn all of it off.

You can see these settings under Settings › Enterprise › Network. An administrator can switch C5 offline there. The proxy settings are delivered by policy or by environment variables, and the card only shows them. Every key can be locked by an organisation policy so nobody on the device can change it.

The platform, on this page, is Growther.si's own service, shown as Mothership in the console — see Connecting to the platform.

Turn off every outbound connection #

One switch stops the connections C5 starts on its own to Growther.si, Hugging Face and the other services listed below. It has more than one name, depending on where you meet it:

  • Egress posture in Settings › Enterprise › Network;
  • Network posture, key networkPosture, values online or offline, in the policy templates and Settings › Enterprise › Policy;
  • "network posture" in some messages.

They are all the same switch, called the offline posture on the rest of this page. To set it:

  1. Open Settings › Enterprise › Network (administrators only).
  2. Under Egress posture, select Offline.

Only an administrator at the machine itself can change this switch: not another user, whatever their permissions, and not from a device paired through Remote Control.

To lock it for a fleet, set it in your policy — see Managed configuration:

yaml
keys:
  networkPosture: { value: offline, locked: true }

Offline stops everything in the table below: the connections C5 makes on its own to Growther.si, Hugging Face, GitHub (for the cloudflared download) or your directory's cloud service, and the connection tests in Settings. There is no second setting to remember for these. Destinations an administrator configured, such as audit shipping and off-site backups, are not stopped; see What offline does not cover.

What What stops
Licence refresh No refresh, no key rotation. Your licence runs on its offline grace period
Platform check-in No check-in
Flywheel sync (the shared catalogue of improvements) Nothing is fetched from the catalogue and nothing is uploaded to it
Background update check Stopped
Memory-search models No embedding, reranking or query-expansion model is downloaded — not at start, and not by growther qmd-run pull
Offline speech model Not downloaded — not at start, and not by growther voice pull
Local classifier model Not downloaded — not at start, not when someone turns the classifier on, and not by growther classifier pull
A classifier model from your providers Asked, and tested, only at an address on this machine. One on your network or online is not asked; the built-in model answers in its place while it is on and Use as fallback is on
Remote Control tunnel program cloudflared is not downloaded. A tunnel that still needs it cannot start; a tunnel whose program is already on the machine still runs (see below)
Hugging Face as a model provider (Inferencer) No background health check (normally at start and once a minute), no model list for the model pickers, and no latency test in Settings, to any Hugging Face host — huggingface.co, hf.co, huggingface.cloud or any name under them, such as router.huggingface.co or a dedicated Inference Endpoint. The provider shows as unhealthy, with no models and no latency figure. An Inferencer endpoint on any other host — your own server — is still checked
Directory group sync Skipped
Sign-in provider test Refused with This deployment's network posture is offline; no outbound probe was made.
Test connectivity Refused with The egress posture is offline, so C5 makes no outbound contact.
growther doctor proxy check Skipped, and the report says so

There is no error-report row: this version of C5 sends no error reports at all, whatever the Report errors anonymously setting says.

Offline is a posture, not a firewall rule: C5 refuses to start the connection itself, so it holds even where the network would have allowed it. Each background job it stops writes one line an hour at most, so you can see the offline posture being honoured without the log filling with it:

text
[posture] offline — local classifier model download skipped; this deployment makes no outbound contact

The name in the middle says which job it was: licence refresh, Mothership check-in, flywheel sync, update check, QMD model download, offline speech model download, local classifier model download, cloudflared download, identity graph sync or identity oidc probe. The Hugging Face provider's skipped checks write nothing at all: a skipped check every minute is not news.

What offline does to the model downloads #

A model download refused by the offline posture is not an error, and nothing is left half done. C5 still checks any model file already on disk, so a model you place by hand is used. Each command tells you what to do instead of downloading:

  • growther classifier pull prints the address to fetch the model from on a connected machine, the exact path to put it at, and its sha256.
  • growther voice pull prints the same three things for the speech model. It first prints Downloading the offline speech model (~60 MB, pinned and sha256-checked)…; the refusal follows straight away, and nothing is downloaded.
  • growther qmd-run pull points you at qmd/README.md in the cache folder, which lists each memory-search model's address, size and sha256.

All three exit with code 1. Settings › Classifier says the same for the local classifier: Not downloaded: This deployment's network posture is offline, so the classifier model is not downloaded., followed by the placement step.

When you switch back to Online, the automatic downloads happen at the next start of C5. You can also run the pull command for the one you want, or — for the local classifier — turn it off and on again in Settings › Classifier, selecting Save after each.

What offline does not cover #

Offline governs what C5 does on its own initiative. It does not stop traffic you have configured: a cloud model provider still answers an agent's request, and a web-search tool still reaches its endpoint, because those happen because someone asked for them. Restrict those separately with the provider allow-list and the guardrail domain allow-list. (C5's own background checks of the Hugging Face provider are stopped, as in the table above; a request an agent sends to a provider is not.)

These connections are not stopped by the offline posture either:

  • A Remote Control tunnel whose program is already on the machine. Offline stops C5 downloading cloudflared, but if the program is already there, the tunnel still starts and connects to Cloudflare's network. Turn Remote Control off to stop it — see Privacy and Cloudflare.
  • LDAP sign-in. C5 still contacts your directory server to sign people in, and Test connection beside the LDAP fields still reaches it. A directory server is normally on your own network, and air-gapped sites are the ones that rely on it.
  • Destinations an administrator configured. Audit shipping to Splunk, Microsoft Sentinel / Azure Monitor, OpenTelemetry or syslog; backups copied to a network folder, an Azure Blob container or an S3 bucket under Settings › Enterprise › Enterprise storage; and scheduled backups to GitHub all keep running while offline. Turn those off where you configured them if the machine must make no outbound contact.

growther update run by hand still contacts the platform. The three model pull commands do not: the offline posture refuses them too. To get a model onto an offline machine, place it by hand — see Place a model by hand.

Details worth knowing #

  • If C5 has not loaded its settings yet when a download is about to start, it uses your policy's value, never "online". A machine set offline by policy makes no connection at all, even while it is starting up.
  • A command you type checks the switch before it downloads anything, but the commands do not all read it the same way:
    • growther classifier pull loads your organisation's whole policy first: the device's own policy (Group Policy, macOS managed preferences or /etc/growther) and a signed policy document, if you use one. Loading the document can mean one request to its web address, or reading the copy C5 keeps of it. A lock set either way wins over the value in Settings. growther classifier status does the same each time it runs while the model is not here and checked, so that it does not suggest a download that the posture or the policy would refuse. Either load happens even while the posture is offline.
    • growther voice pull, growther qmd-run pull and growther doctor read the device's own policy (Group Policy, macOS managed preferences or /etc/growther) and the value saved in Settings. A posture the device policy locks wins over whatever Settings says, so a machine locked offline stays offline for these commands even if someone once saved Online there. Otherwise the value saved in Settings decides; with nothing saved there, they treat the posture as online. An unlocked (Recommended) policy value is not read by these commands, so lock the posture if they must honour it.
    • Those three commands do not read a signed policy document at all. A lock delivered only in a signed document reaches the running C5, growther classifier pull and growther classifier status, but not them. To cover every command, lock the posture through device policy too.
  • On a machine where C5 has never run, these commands use the device policy alone and do not create a database.
  • If C5 cannot read the setting, it keeps the last value it knew. Not being able to read the setting never counts as permission to connect.
  • growther doctor prints Egress posture: offline (zero egress) or Egress posture: online, followed by — locked by managed policy when the device's own policy locks it. While the posture is offline (locked or not), it also skips its proxy check, unless the proxy setting itself cannot be used, which it still reports as a failure.

When an administrator locks this key, the console shows a padlock and the control is frozen.

The models and programs C5 downloads #

Three features run small models on your machine, in their own processes, so your content does not leave it: memory search, voice (offline speech to text) and the local classifier. C5 does not ship those models inside its installer. It downloads each one from Hugging Face the first time the feature needs it, then keeps it. Remote Control's tunnel works the same way with Cloudflare's cloudflared program.

What From When it is downloaded Size Turned off by
Memory-search models (three files) huggingface.co At each start of C5, shortly after it is up, for any of the three that has no file yet about 2.3 GB in total The offline posture, or GROWTHER_QMD_NO_AUTO_DOWNLOAD=1
Offline speech model huggingface.co At each start, while the built-in speech engine is on (it is, unless someone turned it off) and the model is missing about 60 MB The offline posture, or GROWTHER_VOICE_NO_AUTO_DOWNLOAD=1
Local classifier model huggingface.co At start while the classifier is on and the model is missing, and when someone — or a policy — turns it on about 812 MB The offline posture, GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1, a policy that keeps the classifier off, or turning it off
cloudflared (Remote Control tunnel) github.com (Cloudflare's releases) When Remote Control starts a Cloudflare tunnel and the program is missing, or is not the exact file this build pins about 19–69 MB, by platform The offline posture

Each is downloaded once and then reused. A model download that fails is tried again at the next start, not over and over; for the local classifier, turning it off and on, or growther classifier pull, also tries again. A cloudflared download that fails on the network (a dropped connection, an error answer, a download cut short) is retried while the tunnel is on, waiting a little longer each time, up to a minute between tries. One whose contents are wrong — the wrong fingerprint, more than the published size, an archive without the program in it, or a program that fails C5's check — is not retried on its own: the same proxy would serve the same wrong bytes every minute. Remote Control then says The tunnel program downloaded with the wrong contents, so it was not run — a proxy or security product on this network may be altering it. Once that is fixed, save Remote Control's settings to try again. See Troubleshooting Remote Control.

The local classifier's model is never downloaded over a symbolic link placed at its path: a link is yours, typically to a shared copy (see Local classifier).

Some things are true of all of them:

  • Pinned and checked. Every file is fetched from one fixed version — a fixed repository commit on Hugging Face, a fixed release of cloudflared — never from a moving "latest". C5 checks the exact size and the sha256 recorded in the build before it uses the file. A file that does not match is not used.
  • Fetched by C5 itself, through your proxy. Every one of these downloads, whether C5 started it or you typed a pull command, goes through the proxy and CA bundle configured below. The memory-search engine (QMD) cannot download a model itself: C5 blocks it.
  • Nothing sent to Growther.si. A model download is a request from your machine to Hugging Face. Nothing about it is sent to us.

And some are true of the three model downloads:

  • Never on a CI runner. None of them runs automatically when the CI environment variable is set.
  • The memory-search download does not wait for memory search to be used. It runs at start whether or not memory search is paused, so if you never want those 2.3 GB on a machine, set GROWTHER_QMD_NO_AUTO_DOWNLOAD=1 or use the offline posture.
  • The local classifier checks disk room first. Its automatic download does not start unless the free space covers the bytes still to fetch plus a 64 MiB margin (about 67 MB). growther classifier pull does not make that check: it starts the download whatever the free space.

The opt-out variables #

Each one stops only the automatic download for its feature. The matching pull command still downloads, because running it is your explicit request.

Variable Stops Still works
GROWTHER_QMD_NO_AUTO_DOWNLOAD=1 The memory-search models at start growther qmd-run pull, or placing the files by hand
GROWTHER_VOICE_NO_AUTO_DOWNLOAD=1 The speech model at start growther voice pull, or placing the file by hand
GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 The classifier model at start and when it is turned on growther classifier pull, or placing the file by hand

The value must be exactly 1. Anything else, 0 and true included, leaves the download on. Set them where C5's environment is set — the launcher, the service definition, or c5.yaml — see Environment variables.

For an air-gapped machine the offline posture is the better tool: it stops all of them, and the pull commands, with one switch. Where each file goes, and how to place one by hand, is in Where your data lives.

Hugging Face redirects to a download network #

huggingface.co does not serve a model file itself. Its download address answers with a redirect to its download network, and C5 follows it. A proxy or firewall that admits huggingface.co and nothing else therefore fails every model download at the redirect — which usually looks like a download that hangs, or one that fails as soon as it starts. Allow these as well:

  • cas-bridge.xethub.hf.co
  • cdn-lfs.hf.co
  • cdn-lfs-us-1.hf.co
  • cdn-lfs-eu-1.hf.co
  • cdn-lfs.huggingface.co

C5 reaches them only by following that redirect.

github.com does the same for cloudflared: it redirects the download to GitHub's own file-download host, and C5 follows the redirect (never from https down to plain http). Today that host is release-assets.githubusercontent.com; older GitHub setups used objects.githubusercontent.com. The host is GitHub's choice and can change, and C5 does not fix it. If you use Remote Control with a tunnel, allow github.com and that download host: allowing github.com alone fails at the redirect.

If a model download fails #

Nothing breaks while a model is missing: the feature waits for it, and memory search falls back to keyword search. To find out which model is missing and why:

  1. Run growther doctor. Its Local classifier and Offline speech rows say whether each model is here and checked, and if not, why. For memory search, run growther qmd-run doctor.
  2. For the local classifier, Settings › Classifier shows the same under Model.
  3. If the reason is the network, check that huggingface.co and the five download-network hosts above are allowed, through your proxy if you use one.
  4. Run the matching command to try again: growther classifier pull, growther voice pull or growther qmd-run pull.

See also growther doctor.

Corporate proxy #

Setting Policy key Environment variable What it does
Outbound proxy proxyUrl GROWTHER_PROXY_URL An HTTP(S) proxy URL used for every outbound connection
Proxy bypass list proxyNoProxy GROWTHER_NO_PROXY Hosts, domain suffixes and IPv4 CIDRs that go direct
CA bundle caBundlePath GROWTHER_CA_BUNDLE A PEM file of extra certificate authorities, for a TLS-inspecting proxy
Certificate pinning mothershipTlsPin GROWTHER_MOTHERSHIP_TLS_PIN Whether the platform connection pins its certificate. on by default

Loopback always bypasses the proxy, whether or not you list it.

What goes through it. Everything C5 itself sends out: the licence, the platform check-in and Mothership calls, the update check and the self-update download, every model download (the three pull commands included), cloudflared's download, cloud model providers, Test connectivity, and the proxy check in growther doctor. The CA bundle is added to the public authorities, never used in place of them, so hosts your proxy does not re-sign keep working.

C5 does not read HTTP_PROXY, HTTPS_PROXY or NO_PROXY. That is deliberate: a variable in one user's shell must not be able to redirect licence and provider-key traffic. If you want C5 to use the same proxy, set GROWTHER_PROXY_URL (or the policy key) to the same value.

Basic authentication only. C5 can sign in to a proxy with a user name and password in the URL (http://user:password@host:port). NTLM and Kerberos proxies are not supported: give C5 a listener that takes Basic or no authentication, or exclude it from interception.

These four are delivered, not typed. In a shipped build the Network card only shows them. Set them by policy, or by the environment variables in the table (in the launcher environment or c5.yaml); the card is where you confirm what arrived. C5 refuses a change to any of the four from anyone but an administrator at the machine itself: not from another user, whatever their permissions, and not from a device paired through Remote Control. Wherever C5 shows the proxy URL, a password in it is hidden.

An unusable proxy stops traffic; C5 does not go round it. If the proxy setting holds something C5 cannot use — not a URL at all, or a scheme other than http:// or https://, such as socks5:// — C5 does not quietly connect directly instead. Every connection that proxy would carry is refused until the value is fixed or removed; hosts on the bypass list and the machine itself are still reached directly. growther doctor fails its proxy check and says why. The same goes for a proxy URL that your device policy locks but C5 cannot use. C5 also prints a warning once for each such value, naming it, to its own output: [net] the configured proxy … cannot be used: …. That output is the terminal C5 runs in, or its service log where one is kept (see The proxy address stays out of logs and alerts). When C5 starts at sign-in on Windows it is not kept, so growther doctor is the place to look. Its error log gets the same warning without the proxy's address, ending growther doctor names it.

The Network card shows it plainly, whatever the egress posture (going back online does not make a broken proxy usable):

  • the status pill reads Proxy unusable — connections refused instead of Online through a proxy;
  • an alert says C5 refuses to connect through the configured proxy., gives the reason (It cannot be used: …), and says that every connection it would carry is refused, never made directly, until the proxy URL is fixed or removed;
  • the Outbound proxy field shows the value as configured, followed by — cannot be used: <reason>. Any password in it is hidden.

How a policy sets the proxy. The proxy is taken from a policy only when the policy locks it:

  • Locked to a proxy URL — that proxy is used. A locked value C5 cannot use (socks5://, a malformed URL, a list of several blank entries) is refused: every connection the proxy would carry is refused, never made directly.
  • Locked to an empty value (an empty string in policy.d, an empty <string> in a configuration profile, an empty registry value) — no proxy: C5 connects directly, and nobody on the device can set one.
  • Recommended (not locked) — not applied at all, well-formed or not: the value in the environment or c5.yaml decides. A malformed recommended value does not take the machine offline; it is listed among the rejected keys under Settings › Enterprise › Policy.

C5 gives the same answer from the moment it starts, before it has loaded your whole policy, as after.

A change to any of the four is picked up by a running C5 within about 30 seconds: the outbound stack is rebuilt in place, with no restart. Connections already open finish on the settings they started with.

The console may still show a "restart needed" hint for these four. You can ignore it: C5 picks the change up within about 30 seconds.

These four can come only from device policy (Group Policy, macOS managed preferences or /etc/growther) or from the environment, never from a signed policy document on a share or web address. Otherwise, a document that changed your proxy could also intercept the download of the next document.

The certificate pin and your proxy #

With mothershipTlsPin on, platform calls still go through your proxy. The proxy opens a tunnel, and the pin is checked on the encrypted connection inside it — so a proxy that only passes traffic through works with the pin on. Only a proxy that decrypts and re-signs traffic needs the pin off, as below.

Behind a TLS-inspecting proxy #

If your proxy re-signs TLS, C5 will not trust it until you give it the authority. Deliver both values the same way you deliver the rest of your policy:

  1. Export your proxy's root CA as PEM and put it somewhere every device can read.
  2. Set caBundlePath to that path.
  3. Set mothershipTlsPin to off — the pin exists to detect exactly what your proxy is doing, so switching it off must be a deliberate, recorded decision rather than a silent failure.
  4. Wait about 30 seconds, or restart C5 if you would rather not wait.
  5. Check it: select Test connectivity on the same screen, or run growther doctor and look at the Proxy reachability row. Every host should answer.

Turning pinning off is the one step here that removes a protection. Leave it on unless you are actually behind an inspecting proxy.

When the proxy says no #

A failed update check, a failed growther update, or a failed model download (the local classifier's, the offline speech model's or memory search's, whether C5 started it or you ran a pull command) names the cause rather than only saying "fetch failed": the certificate pin, an untrusted certificate, or the proxy that refused. The local classifier's message gives the cause straight after Could not download the classifier model:; the memory-search and speech downloads give it after fetch failed:. In a terminal — the output of growther update, a pull command or growther doctor, and memory search's line in C5's own output — a proxy refusal names the proxy:

text
the proxy http://proxy.corp.example:3128 refused to open a tunnel to api.growther.si:443: it answered CONNECT with 407 Proxy Authentication Required; it wants credentials — put them in the proxy URL (http://user:password@host:port), C5 speaks Basic proxy authentication only

Everywhere else — in Settings, in C5's error log, in notifications — the same sentence says the proxy, without its address (see The proxy address stays out of logs and alerts):

text
the proxy refused to open a tunnel to api.growther.si:443: it answered CONNECT with 407 Proxy Authentication Required; it wants credentials — put them in the proxy URL (http://user:password@host:port), C5 speaks Basic proxy authentication only

Credentials in your proxy URL are never printed. The end of the line tells you what to do:

What the message says What it means, and what to do
The proxy answered CONNECT with 407 The proxy wants credentials. Put them in the proxy URL (Basic authentication only)
The proxy answered CONNECT with 403 Its policy does not allow this host. Ask for the host to be allowed
The proxy answered CONNECT with 502, 503 or 504 The proxy could not reach the host itself
Mothership TLS public-key pin mismatch Your proxy is decrypting and re-signing the platform connection. Follow Behind a TLS-inspecting proxy above
A certificate C5 does not trust, in the system's own words — for example self-signed certificate in certificate chain or unable to get local issuer certificate. Through the proxy, outside a terminal: the certificate presented on the way to … is not trusted (…) Your proxy re-signs traffic with its own authority. Add its root CA with caBundlePath
C5 refused to connect to … the configured proxy … cannot be used The proxy setting is not something C5 can use (see An unusable proxy stops traffic above). Fix or remove it

growther doctor checks the proxy from the command line. With a proxy configured, its Proxy reachability row asks the proxy for four hosts: the platform, the licence service, the release mirror and github.com. It asks no more than four, so if the proxy is down, doctor waits for one round of time-outs, not many. It warns with every host that failed: Allow these hosts on the proxy, or add its CA to the CA bundle. If the proxy setting is one C5 cannot use, the row fails without asking any host, and says the setting must be an http:// or https:// address.

The proxy address stays out of logs and alerts #

Your proxy's address — its host name, IP address and port, and any user name or password in its URL — describes your network, so C5 shows it only where an administrator works. It is kept out of these places, which anyone signed in to C5 can read:

  • C5's error log (logs/errors.log in your C5 folder), and the error list and alerts on the Monitor page. C5 also sends each new line to every open window, paired phones included;
  • notifications, the New Error Cluster Detected one included, and the emails C5 sends for them;
  • the grouped errors under Ops › Errors and in the Errors tab of the HITL window, and the notification rows of the audit log, on screen and in Export CSV;
  • the status C5 shows for a download: Remote Control's tunnel line, the Model line of the built-in model on the Classifier card, and the progress of an update started from the app;
  • the error text C5 gives a model when it triages repeating errors and notifications.

These say the proxy, and what it answered and for which host — for example the proxy refused to open a tunnel to github.com:443: it answered CONNECT with 407 …, or the connection to huggingface.co:443 through the proxy failed (ECONNREFUSED). A proxy setting C5 cannot use reads the configured proxy cannot be used, followed by growther doctor says why or by the reason.

Older text. Log lines, notifications and error groups written by v322 or earlier, before C5 kept the proxy out of them, or by an older version after a rollback, are cleaned when C5 shows them. C5 cannot know which proxy an old line refers to, so in those lines it hides every network address in a connection error except this machine's own, and shows (address withheld). The log file itself is not rewritten: older lines still contain the address on disk, for anyone who can open logs/errors.log on the machine.

Where an administrator sees which proxy failed:

  • growther doctor on the machine. Its Proxy reachability row starts with the proxy in force, and its Mothership row names the proxy that refused;
  • C5's own output: the terminal it runs in, or, when C5 starts at sign-in, logs/service.out.log and logs/service.err.log in your C5 folder (macOS) or journalctl --user -u growther-c5 (Linux with systemd). It carries the [net] the configured proxy … cannot be used warning, and memory search's line when its download fails. When C5 starts at sign-in on Windows, or from a Linux desktop's autostart, this output is not kept, so use growther doctor instead;
  • Settings › Enterprise › Network, whose Outbound proxy field shows it with any password hidden;
  • in a terminal, growther update, growther classifier pull, growther voice pull and growther qmd-run pull, which name the proxy that refused, as in the first example above.

What C5 contacts, and why #

This is the list for a firewall ticket. Test connectivity on the same screen (administrators only) asks each host on it through your real proxy settings and reports which answered. growther doctor prints the same list on its Control-plane hosts row.

Host Why
api.growther.si The signed version manifest and update catalogue, the licence check-in clock, and the shared catalogue of improvements (flywheel sync)
license.growther.si Device-code activation, licence refresh and key rotation. C5 checks the licence's signature itself, so this host only carries it
raw.githubusercontent.com The release mirror: installer scripts and self-update binary assets. Every asset is SHA-256 checked against the signed manifest
github.com Remote Control: C5 downloads Cloudflare's cloudflared from its releases there, only when a Cloudflare tunnel starts and the program is not already on the machine. Its size and SHA-256 are checked before it runs
release-assets.githubusercontent.com Where github.com redirects that download today. Reached only by following that redirect. The host is GitHub's choice and can change (above)
huggingface.co Model downloads: the three memory-search models, the offline speech model, and the local classifier's model (only while the classifier is on, or when someone runs growther classifier pull). Stopped by the offline posture; each automatic download also by its GROWTHER_*_NO_AUTO_DOWNLOAD variable
cas-bridge.xethub.hf.co Hugging Face's download network. Reached only by following a redirect from huggingface.co
cdn-lfs.hf.co Hugging Face's download network, as above
cdn-lfs-us-1.hf.co Hugging Face's download network (US), as above
cdn-lfs-eu-1.hf.co Hugging Face's download network (EU), as above
cdn-lfs.huggingface.co Hugging Face's older download-network name, as above
api-inference.huggingface.co The default Hugging Face inference endpoint (the Inferencer provider) — only when you have configured that provider with an API token and no address of your own. An HF_TOKEN or INFERENCER_API_KEY in C5's environment counts as one. An install that has not configured this provider sends it nothing, and while the posture is offline C5 makes no background check of it
growther.si The canonical installer scripts, and product documentation linked from the console
docs.growther.si Documentation, reached only when someone clicks a link

That is fourteen hosts by default; growther doctor prints the same count. The two GitHub hosts matter only if you use Remote Control with a Cloudflare tunnel, but allow both if you do: allowing github.com alone fails at the redirect. Once the tunnel runs, it connects to Cloudflare's network — see Privacy and Cloudflare.

If you point C5 at a relay or mirror with Platform address (GROWTHER_PLATFORM_BASE_URL), that host is listed first and the default is still listed — a binary that has not yet read your policy, such as the installer, will use the default.

Nothing on this list carries your work. The databases, your notes and your deliverables never leave the machine.

Why C5 listens on two addresses #

localhost is not one address. It is two — 127.0.0.1 on IPv4 and ::1 on IPv6 — and which one a program gets depends on how its resolver is configured. C5 serves its own address bar name, and passkeys are bound to it, so it listens on both by default.

You will see both in the boot log:

text
[bootstrap] C5 API running on http://127.0.0.1:4299
[bootstrap] C5 API also on http://[::1]:4299

Open C5 at http://localhost:4299 all the same; that is the address C5 opens in your browser. Passkeys are tied to the name localhost, not to either number.

This changes nothing about who can reach C5. Both addresses are loopback, so the server is still reachable only from the computer it runs on, and the host check, the browser-origin allowlist and the passkey origin rules all behave exactly as they did.

If IPv6 is switched off on your machine — some hardened Linux builds and some Windows configurations do this — the second listener cannot start. That is not a failure and C5 does not stop for it; you get a line naming the reason and the address that does work:

text
[bootstrap] no ::1 listener (EAFNOSUPPORT) — C5 is reachable on 127.0.0.1 only. If a browser resolves localhost to ::1 it will need http://127.0.0.1:4299 instead.

Setting GROWTHER_BIND_HOST to a real interface or to 0.0.0.0 is a deliberate choice to widen exposure, and C5 does not pair those with a second address: you asked for one family, and pairing would quietly add the other.

Making C5 reachable from another machine #

By default C5 binds to loopback only: the server is reachable from the computer it runs on and from nowhere else. To serve it to other machines you must say so explicitly, and say which names you will serve.

Variable Default What it does
GROWTHER_BIND_HOST 127.0.0.1 The interface to bind. Set 0.0.0.0 for a deliberate self-host deployment
GROWTHER_ALLOWED_HOSTS loopback names Comma-separated host names C5 will answer to
GROWTHER_ALLOWED_ORIGINS the C5 origin Comma-separated browser origins allowed to call the API. Never a wildcard
GROWTHER_WEBAUTHN_ORIGINS the C5 origin Origins a passkey may be registered and used on
PORT 4299 The port C5 serves on, in every environment

Three things to get right together, or sign-in breaks in a way that is hard to read:

  • Binding to a real interface turns the host check off unless you turn it back on. Bound to loopback, only loopback host names are accepted. Bound to 0.0.0.0, setting GROWTHER_ALLOWED_HOSTS enables the check for exactly the names you list — and leaving it unset is an explicit opt-in to answering any Host header at all. Set it.
  • WebAuthn origin checking is exact. A passkey collected on https://c5.corp.example is refused at https://c5.corp.example:8443. If you serve C5 under a real name, set GROWTHER_WEBAUTHN_ORIGINS to match exactly what the browser will show — including the scheme and any port.
  • PORT moves everything. The allowed origins and the WebAuthn defaults are derived from it, so changing the port does not strand them — but a hard-coded origin in your own reverse-proxy configuration will.

Serving C5 beyond loopback also means the loopback trust boundary no longer does the work it used to. Put it behind a reverse proxy that terminates TLS, and read Reaching C5 by a corporate name, which covers moving passkeys to that name without invalidating the ones people already hold.

Edit this page on GitHub ↗Last updated Nov 20, 2026