Skip to contentdocs.Growther.si
Download C5

Managed configuration

C5 can take its settings from your organisation instead of from each user. Push a policy from the tools you already run, and every setting it locks shows Managed by your organization on the user's screen.

This page is for IT administrators. Users need to do nothing.

What you can manage #

Every key an organisation may set is listed in the generated templates (below) and under Settings › Enterprise › Policy on any C5. In short:

  • Where things live: the configuration, data, secrets and cache folders (device policy only).
  • Network: port, bind address, allowed hosts, whether credential-less sign-in is allowed from the network, the proxy and the certificate authority. The proxy and certificate keys come from device policy only, never from a signed document, and only when the policy locks them — see How a policy sets the proxy below.
  • Offline mode: networkPosture (online or offline), the one switch that stops the outbound connections C5 starts on its own (not the audit shipping or off-site backups you configure), for air-gapped fleets — see Turn off every outbound connection. Lock it through device policy: a device-policy lock is honoured by every growther command, even on a machine whose Settings once said Online, while a lock that arrives only in a signed document is not read by growther voice pull, growther qmd-run pull or growther doctor (see Details worth knowing).
  • What leaves the machine: sharing improvements with the Mothership, anonymous error reports (the key exists, but this version sends no error reports — see Network and egress), GitHub backups, automatic updates and the release channel.
  • Models: which providers and models agents may use, whether C5 runs its local classifier (classifier_enabled), how far away a classifier model from your providers may run (classifier_external_allowed), and whether the built-in model answers in its place when it cannot be reached (classifier_builtin_fallback). All three are described below.
  • Security: elevated exec, workspace-only edits, autonomy ceilings, the global cost cap, allowed domains.
  • Retention and audit: audit and event retention, legal hold, the audit shipping endpoint.
  • Identity: the sign-in provider, admin role, whether single sign-on is required.

Secrets are never part of a policy. API keys and tokens live in the keystore; a policy refers to them, it never carries them.

Four ways to deliver a policy #

First: get the templates #

C5 writes them itself. Every install puts the current set in <home>/enterprise at startup — %USERPROFILE%\.growther\enterprise on Windows, ~/.growther/enterprise elsewhere — and you can write them anywhere on demand:

bash
growther policy templates --out ./c5-mdm
File For
Growther-C5.admx + en-US/Growther-C5.adml Group Policy and Intune
ai.growther.c5.mobileconfig Jamf, Intune and any macOS MDM
samples/policy.d/10-baseline.yaml Linux drop-in, ready to edit
samples/policy.example.yaml A signed policy document to start from
growther-c5-policy.schema.json Validate a document in your editor or CI
README.md Every key, its type and what it locks

Use the templates from the version you are deploying. They are generated from that build's own key list, so they can never offer a key the server does not understand — which is the failure they exist to prevent: a policy an administrator sets and the server silently ignores. C5 refreshes them on every start, and leaves any file you have edited alone. Templates written by a build from before the local classifier do not carry classifier_enabled (below), and those from before C5 v327 (Terms of Service v1.9) do not carry classifier_external_allowed (below), and those from before Terms of Service v1.10 do not carry classifier_builtin_fallback (below); write them again from the version you deploy.

Windows: Intune or Group Policy #

Import Growther-C5.admx and en-US/Growther-C5.adml into the Intune Settings Catalog (imported ADMX) or your central store. The policies are under Administrative Templates › Growther › C5, in both Computer Configuration and User Configuration:

  • Computer Configuration locks. The value lands in HKLM\SOFTWARE\Policies\Growther\C5, and users see the setting disabled with a lock; a change made in C5 is ignored and audited.
  • The Recommended twins, under C5 (Recommended: values that do not lock), do not lock. Their values land in …\C5\Recommended. Templates written by earlier versions titled this folder C5 (Recommended: defaults the user may change), which was not true of the eight keys below.
  • User Configuration never locks. A policy set there lands under HKCU, and C5 reads it as it reads a Recommended value.
  • Disabled on a policy that holds a value (a folder, a port, a choice from a list) works like Not Configured: nothing is written, so C5 behaves as if the policy were not set. On an on/off policy, Enabled sets it to true and Disabled to false.

A value that does not lock — Recommended, or under User Configuration — is for most keys a default: C5 uses it while the setting has no value of its own in C5. Eight keys apply only when locked, and a value that does not lock does nothing for them: GROWTHER_HOME, GROWTHER_DATA_DIR, GROWTHER_SECRETS_DIR, GROWTHER_CACHE_DIR, PORT, GROWTHER_BIND_HOST, GROWTHER_ALLOW_REMOTE_NONE_LOGIN and GROWTHER_ALLOWED_HOSTS. The proxy and certificate keys have rules of their own; see How a policy sets the proxy. Each policy's explanation in the templates says which applies to it.

macOS: Jamf or Intune #

Deploy a configuration profile for the preference domain ai.growther.c5. Top-level keys are locked; keys inside a Recommended dictionary are defaults, except the keys the profile marks lock only (the eight above), which a Recommended value does not set. A top-level key in a profile for one user still locks, but for that user only, and it never lifts a lock on classifier_enabled (see With classifier_enabled locked off). Use the ai.growther.c5.mobileconfig written above as your starting point.

Linux #

Drop one or more YAML files in /etc/growther/policy.d/. They are merged in name order and use the same document shape as a signed policy file. Start from samples/policy.d/10-baseline.yaml.

That directory is root-owned on purpose, and so is the signing keyring (/etc/growther/policy-keys on Linux, %ProgramData%\Growther\C5\policy-keys on Windows, /Library/Application Support/Growther/policy-keys on macOS). C5 never reads policy from the user's home folder: if it did, the person the policy governs could write one. The templates in <home>/enterprise are copies to export — nothing there influences what C5 believes.

A signed policy file (any platform) #

Publish a policy document on a share or at an https address and point devices at it through the platform policy (PolicySource), or through /etc/growther/policy-source.yaml on Linux. File and https sources must be signed with a key you pin on each device:

bash
growther policy keygen --out ./policy-keys          # once, on an admin workstation
growther policy sign policy.yaml --key ./policy-keys/policy-signing.key --out policy.signed.yaml
sudo growther policy pin ./policy-keys/policy-signing.pub   # on each device, or through your MDM

A policy document looks like this:

yaml
schemaVersion: 1
policyId: contoso-c5-baseline
version: 12
managed: true
maxOfflineHours: 336
keys:
  shareImprovements: { value: false, locked: true }
  allowedProviders: { value: ["anthropic", "azure-openai"], locked: true }
  auditRetentionDays: { value: 730 }

Azure App Configuration and AWS AppConfig are also accepted as sources; their identity stands in for the signature, and a signature is verified when present.

How a policy is applied #

  1. C5 reads the device policy, then the signed document, at start and every fifteen minutes (or on demand).
  2. The signature is checked against the pinned key, and the version must be higher than the last one applied. An older document is refused as a replay.
  3. Every key is checked on its own. An unknown key, a secret, or a value of the wrong type is reported and skipped. A locked key with a bad value falls to its safest default and raises a critical audit event. One typo never unlocks a fleet.
  4. Locked keys shadow whatever the user had set. Lifting the policy restores the user's value. A teammate's save that changes a locked setting is refused, nothing in it is stored, and the message names the setting as the policy templates do, for example Forbidden: Share improvements with the Mothership is managed by your organisation's policy. An administrator's save goes through without the locked settings, and on Settings › Budget and Settings › Remote Control a message names the ones left out as the tab does: Not saved — managed by your organisation's policy: Hard Budget Cap spend limit., or Policy on this machine refused: Read-only for remote devices.
  5. If the source becomes unreachable, the last verified policy holds for the grace window (default fourteen days). After that, locked keys fall to their safest defaults and C5 says so. It never quietly runs unmanaged.
  6. To stop managing on purpose, publish a signed document with managed: false.

How a policy sets the proxy #

proxyUrl behaves differently from most keys, because a proxy decides where every connection C5 makes is sent:

What the device policy says What C5 does
Locked to a usable http:// or https:// URL Uses that proxy
Locked to an empty value (proxyUrl: { value: "", locked: true } in policy.d, an empty <string> in the profile, an empty registry value) No proxy: connects directly, and nobody on the device can set one
Locked to a value C5 cannot use (socks5://…, a malformed URL, a list of several blank entries) Refuses every connection the proxy would carry, and never connects directly instead. The Network card reads Proxy unusable — connections refused
Recommended (not locked), well-formed or not Not applied. The proxy in C5's environment or c5.yaml decides. A malformed recommended value is only listed among the rejected keys, so a typo in it cannot take a machine offline

C5 reads the same answer when it starts, before your whole policy has loaded, as it does afterwards. See Corporate proxy.

What users see #

A locked control is disabled, shows the value your policy set, and carries a small lock that names the source when hovered. Administrators who save a page that includes a locked setting see which settings were set aside; nothing else on the page is lost.

Settings › Enterprise › Policy shows the source, the policy id and version, the signature state, when it was fetched and applied, the locked settings, the defaults it applied, and any keys it rejected with the reason.

Keep the local classifier on or off #

C5's local classifier is a small model (about 812 MB, downloaded once from Hugging Face) that answers some of C5's routine questions on the machine itself. It is on by default: the first time a version of C5 with the classifier starts, it turns the switch on wherever nothing has set it, and then downloads the model. While it is on, every kind of decision starts in Shadow, where it only records its answers. Passing its measurement changes nothing by itself: only Task categories can be raised to act, and only when someone with permission to change settings chooses to. No policy key does it. On a machine where it runs on the CPU, Step quality records one question in four; every other kind is recorded every time. What it does, and what users see, is described in Classifier and under Settings › Classifier, where its switch is Built-in model. One policy key decides it for a fleet:

Key classifier_enabled
Type Boolean: true keeps it on, false keeps it off
Lockable Yes
Delivered by Group Policy or Intune, macOS managed preferences, /etc/growther/policy.d, or a signed policy document. Not by the Mothership
Applies Live, with no restart
If a locked value is malformed, or the policy expires Off

In the generated templates it is the Classifier policy in Growther-C5.admx (enabled is 1, disabled is 0, under the registry value classifier_enabled), and a commented-out classifier_enabled line in ai.growther.c5.mobileconfig and in the YAML samples — uncomment it and set the value. To keep it off everywhere:

yaml
keys:
  classifier_enabled: { value: false, locked: true }

Set it before the first start if you never want the download #

A policy value counts as an answer. If classifier_enabled arrives before the first start of a C5 version with the classifier, C5 never turns the switch on itself, so a fleet that sets false never downloads the model at all. The notice C5 normally gives administrators — The local classifier is on — is not sent either.

A Recommended (unlocked) value is only a default. It decides machines where nobody has set the switch yet; on a machine that has already started such a version, C5 has already set it on, and that setting outranks a recommended false. To turn it off on machines that are already running, lock it.

What a lock does #

A change takes effect at the next policy reload — at most fifteen minutes, growther policy reload, or Reload policy in Settings › Enterprise › Policy — with no restart:

Locked off. A running classifier is stopped, and a download in progress is stopped too (the part already downloaded stays, so it can resume if the policy changes). Nothing is downloaded afterwards: not at start, not from Settings, and not by growther classifier pull, which answers Your organisation's policy keeps the local classifier off, so its model is not downloaded. and exits with code 1. While the model is not on the machine, growther classifier status says so too, rather than suggesting a pull: model not downloaded: your organisation's policy keeps the local classifier off. Both read a lock in a signed policy document as well as one in the device's own policy. A model already on the machine stays there until an administrator deletes it; the Delete the model button in Settings › Classifier still works while the classifier is off.

A lock to false also keeps off every classifier model from your providers, unless your policy names classifier_external_allowed at the right level — see With classifier_enabled locked off.

Locked on. Nobody on the device can turn it off or delete its model, because C5 would only download it again. A request to turn it off is refused with Locked by your organisation's policy: classifier_enabled., and a delete with Your organisation's policy keeps the local classifier on, so its model was not deleted: C5 would download it again. The offline posture and the GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD variable still stop the download itself — see Network and egress. If the file at the model's path is not the model, the card says so and that Your organisation's policy keeps the classifier on, so the file cannot be deleted here. An administrator at the machine can put the model in its place by hand.

In Settings › Classifier, either way, the Built-in model switch is shown disabled with the managed badge, and the card says why in words a phone user can read without hovering: Your organisation's policy keeps the local classifier off, so it cannot be turned on here. The policy is set by … followed by where the policy came from. Where the lock keeps every model off, the line reads instead: Your organisation's policy keeps the local classifier off, so no model answers these questions here, built-in or from your providers. The Terms dialog does not offer Turn off the local classifier…, and under a lock that keeps it on, no delete button is shown.

Lifting the lock gives the switch back to the device: it returns to whatever was set there, and the classifier starts or stops to match. On a machine where nothing was ever set (the policy arrived before the first start), it stays off until C5 next starts, and C5 then turns it on as it would on a new install.

Classifier models from your providers #

An administrator can also choose a model that one of your model providers serves to answer the classifier's questions — see A model from your providers. None is chosen by default. One key decides how far away such a model may run:

Key classifier_external_allowed
Values off (never), loopback (only a model C5 can show runs on this machine), lan (this machine or your network), any (anywhere, a provider's cloud included)
Lockable Yes
Delivered by Group Policy or Intune, macOS managed preferences, /etc/growther/policy.d, or a signed policy document. Not by the Mothership
Applies Live, with no restart
When no policy sets it An administrator at the machine chooses it in C5. Until they do, loopback
If a locked value is malformed, or the policy expires off

In C5 this is Models allowed from, in Settings › Classifier: This machine (loopback), Your network (lan) or Anywhere (any). Only an administrator at the machine can change it there — never from a device paired through Remote Control.

Who decides #

Highest first:

  1. A policy that locks the key. The card shows Models allowed from locked, with the reason and where the policy is set (for off, the choice reads Off — set by your organisation). Nobody in C5 can change it.
  2. What an administrator at the machine chose on the card.
  3. A value that does not lock: a Recommended value, or one set under User Configuration. It is a default the administrator may change on the card. A recommended off also shows Off — set by your organisation, but there it can be changed: choosing This machine, Your network or Anywhere replaces it. Nobody can choose off on the card itself.
  4. Nothing set: loopback, shown as This machine.

What each value lets through #

Where a model runs is what C5 can show, not what its address suggests:

  • loopback (This machine): a llama.cpp server or an Ollama model on this machine, when that program itself tells C5 it runs the model here. An address on this machine is not enough on its own: a program there can pass the questions on to another computer. A llama.cpp server that hands part of its model to RPC servers on other computers can still say it runs here: a single-model server started with --rpc, or a router's model given LLAMA_ARG_RPC, shows it nowhere C5 can read (see Who answers).
  • lan (Your network): also a model on another computer on your network, and LM Studio or a vLLM server on this machine. Neither can show C5 where its model runs; LM Studio, for one, can serve a model that runs on another device. So C5 counts them as your network even at this machine's address.
  • any (Anywhere): also the cloud providers' models (OpenAI, MS Azure, OpenRouter, DeepSeek, xAI Grok, Alibaba Cloud Qwen, Moonshot Kimi, Meta and Google Gemini). Those providers bill your account for each question.

Whatever the value, two things hold:

  • A model off this machine still needs an administrator's confirmation. C5 sends a model on your network or online none of your work until an administrator at the machine has seen where it runs and what C5 would send it, and confirmed. Widening the limit does not skip that.
  • Tool-call safety questions are asked only of a model C5 can show runs here. The one gap is the llama.cpp --rpc offload above: such a model counts as this machine and is asked them too. Do not let a model that is offloaded that way be chosen, or keep its RPC servers on the same machine.

allowedProviders and allowedModels apply too: a provider or model they do not allow cannot be chosen for the classifier either.

Narrowing it takes effect at the next policy reload. A model chosen earlier that now falls outside the limit stops being asked. Its panel on the card reads Not used., followed by the reason, and the built-in model answers in its place while it is switched on and Use as fallback is on (below).

In the generated templates it is the Classifier models from your providers policy in Growther-C5.admx, a drop-down of off, loopback, lan and any (the string registry value classifier_external_allowed), with a (recommended) twin. Set to Disabled, it behaves as if it were not configured. It is also a commented-out line in ai.growther.c5.mobileconfig and in the YAML samples. To forbid every model from your providers:

yaml
keys:
  classifier_external_allowed: { value: "off", locked: true }

classifier_builtin_fallback #

While a model from your providers is chosen, the built-in model normally answers in its place whenever that model cannot be reached. This key can stop that. In C5 it is Use as fallback, under the built-in model's switch on the Classifier card (see Use as fallback).

Key classifier_builtin_fallback
Values true (the built-in model answers in the chosen model's place) or false (nothing does, and C5 uses its usual checks)
Lockable Yes
Delivered by Group Policy or Intune, macOS managed preferences, /etc/growther/policy.d, or a signed policy document. Not by the Mothership
Applies Live, with no restart
When no policy sets it What was chosen on the card. Until someone does, true
A Recommended value, or one set under User Configuration A default, which someone allowed to change it on the card may change
If a locked value is malformed false

false keeps the built-in model out of the chosen model's place only. It still answers what the chosen model is not asked, such as Tool-call safety beside a model off this machine, and still acts where it has a promotion record of its own and the chosen model has none. To keep the built-in model from answering anything, lock classifier_enabled to false as well.

The key matters only while a model from your providers is chosen and turned on, and the built-in model is on. Builds that ship Terms of Service earlier than v1.10 ignore it, and the built-in model answers in the chosen model's place there.

Without a policy, anyone who may change settings in C5 can turn Use as fallback off, from a paired device too; turning it back on is for an administrator at the machine. Locked, the switch is greyed out, with Your organisation's policy sets this. The policy is set by … under it.

In the generated templates it is the Classifier built-in model as fallback policy in Growther-C5.admx: Enabled writes 1 (on) and Disabled writes 0 (off). It has a (recommended) twin, and is a commented-out line in ai.growther.c5.mobileconfig (<key>classifier_builtin_fallback</key><false/>) and in the YAML samples. To keep the built-in model out of a chosen model's place on every machine:

yaml
keys:
  classifier_builtin_fallback: { value: false, locked: true }

With classifier_enabled locked off #

Locked to false, classifier_enabled keeps every classifier model off: the built-in one and any from your providers. The one exception is a policy that also sets classifier_external_allowed, locked or only recommended:

  • at the same level as the lock, or a higher one. Device policy (Group Policy or Intune, macOS managed preferences, /etc/growther/policy.d) is a higher level than a signed policy document;
  • and for the computer. A value set for a user — under User Configuration, or in a configuration profile for one user — never lifts the lock.

What an administrator chooses on the card never lifts it either. Set like that, models from your providers may answer within the limit while the built-in model stays off and is never downloaded. For example, to allow only models that run on each machine:

yaml
keys:
  classifier_enabled: { value: false, locked: true }
  classifier_external_allowed: { value: "loopback", locked: true }

While the lock keeps every model off, the card says so once for both panels: Your organisation's policy keeps the local classifier off, so no model answers these questions here, built-in or from your providers., followed by where the policy is set.

Validate before you ship #

bash
growther policy validate policy.signed.yaml
growther policy show
growther policy reload

validate prints one row per key and exits non-zero if anything would be rejected.

Unattended installs #

To push C5 itself, use the Windows or macOS installer package — see Deploying across an organisation.

Add --policy-expected to the install script, or run growther policy expect after installing. C5 then treats the absence of a policy as a fault rather than a choice, and disables the anonymous first-user setup so nobody can claim the administrator seat before your policy arrives. Create the first administrator from the device:

bash
growther user bootstrap-admin --name "IT Admin"

or let the first sign-in through your identity provider create it.

growther doctor --json reports Managed install with the marker and the admin count, which is what an Intune detection script or a Jamf extension attribute should read.

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