Environment variables
Environment variables let you change how C5 behaves without editing any files. You set them before the command, like this:
PORT=5000 growtherOr you can set them for your whole terminal session:
export PORT=5000
growtherIf C5 starts on its own at sign-in, a variable you set in a terminal does not reach it. See When C5 starts at sign-in below.
The variables #
Where things live
| Variable | What it does | Default |
|---|---|---|
GROWTHER_HOME |
Where C5 keeps all your data. | ~/.growther |
GROWTHER_DATA_DIR |
Where the databases live. Must be a local disk. | <home>/data |
GROWTHER_SECRETS_DIR |
Where the device key set lives. Local only. | <home>/config |
GROWTHER_CACHE_DIR |
Where downloaded models and caches live. | <home> |
Running and updating
| Variable | What it does | Default |
|---|---|---|
PORT |
Which port the C5 server listens on. | 4299 |
GROWTHER_CHANNEL |
Which releases you get: stable, beta, or dev. |
stable |
GROWTHER_NO_SELF_INSTALL |
Set to 1 to run the program where it sits. |
off |
GROWTHER_NO_MODIFY_PATH |
Set to 1 so C5 never edits your shell files. |
off |
GROWTHER_UPDATE_RESTART |
Set to exactly 0 (or pass --only) so C5 does not restart or ask to start after an update. See growther update. |
restarts if running, asks to start if stopped |
GROWTHER_ROLLBACK_FORCE |
Set to 1 to roll back past the keystore check, like --force. |
off |
Model downloads
| Variable | What it does | Default |
|---|---|---|
GROWTHER_QMD_NO_AUTO_DOWNLOAD |
Set to 1 so C5 never downloads the memory-search models on its own. |
off |
GROWTHER_VOICE_NO_AUTO_DOWNLOAD |
Set to 1 so C5 never downloads the offline speech model on its own. |
off |
GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD |
Set to 1 so C5 never downloads the local classifier's model on its own. |
off |
Network
| Variable | What it does |
|---|---|
GROWTHER_PROXY_URL |
The HTTP(S) proxy for every outbound connection. |
GROWTHER_NO_PROXY |
Hosts, domain suffixes and IPv4 ranges that skip the proxy. |
GROWTHER_CA_BUNDLE |
A PEM file of extra certificate authorities. |
GROWTHER_MOTHERSHIP_TLS_PIN |
on or off: whether the platform connection pins its certificate. |
These four are explained in Network and egress. C5 does
not read HTTPS_PROXY, HTTP_PROXY or NO_PROXY; set GROWTHER_PROXY_URL to the
same value if that is the proxy you want.
Local model servers and gateways
| Variable | What it does |
|---|---|
LMSTUDIO_BASE_URL |
Where LM Studio is. |
OLLAMA_BASE_URL |
Where Ollama is. |
LLAMACPP_BASE_URL |
Where llama.cpp's server is. |
INFERENCER_BASE_URL |
Where the Inferencer (Hugging Face) endpoint is. |
OPENCLAW_GATEWAY_URL |
Where the OpenClaw gateway is. |
HERMES_GATEWAY_URL |
Where the Hermes gateway is. |
HERMES_DASHBOARD_URL |
Where the Hermes dashboard is. |
PICOCLAW_GATEWAY_URL |
Where the PicoClaw gateway is. |
See Local model servers and gateways below.
Image library
| Variable | What it does |
|---|---|
GROWTHER_LIBVIPS_DIR |
Run C5 on your own copy of libvips instead of the one it ships with. |
Troubleshooting
| Variable | What it does |
|---|---|
GROWTHER_DEBUG_CATEGORIZATION |
Set to 1, true, yes or on to log each task categorisation to logs/categorization-debug.log; anything else, 0 and false included, leaves it off. The log names a goal by a fingerprint and its length, and keeps only the first 300 characters of the model's reply, with secrets and personal details masked; the reply can still repeat words from the task. The same switch, and the same masking and cut, decide whether C5's own output (the terminal, or logs/service.out.log at sign-in on macOS) shows the model's reply to a categorisation or to a task's title; off, it shows at most their lengths. errors.log never holds the reply. Leave it off unless you are asked for it. |
Media tools
| Variable | What it does |
|---|---|
FFMPEG_PATH |
The ffmpeg program render_video uses. Looked at after the FFmpeg location in Settings › Tools; then ffmpeg on your PATH, /opt/homebrew/bin, /usr/local/bin, /opt/local/bin and /usr/bin. See Pictures and video. |
Common things people do #
Use a different port #
If something else on your computer already uses port 4299:
PORT=5000 growtherKeep your data somewhere else #
Handy if you want your C5 data on an external drive, or you want two separate setups that do not share anything:
GROWTHER_HOME=/Volumes/Work/growther growtherWarning Each data folder is its own separate world. Chats, tasks, and settings in one folder are not visible from another.
Put the downloaded models on another disk #
The models C5 downloads take about 3.1 GB in all if you use every feature. They are not your data and are never backed up, so they can live on a bigger, slower disk:
GROWTHER_CACHE_DIR=/Volumes/Big/c5-cache growtherC5 decides where the cache lives in this order, and the first answer wins:
- A
cacheDirset by your organisation's policy. GROWTHER_CACHE_DIR.- The location recorded by
growther cache relocate. - Your C5 folder.
For a permanent move, growther cache relocate --to <folder> is usually better: it moves
what is already downloaded and records the location for you. It refuses while
GROWTHER_CACHE_DIR is set, because the variable would win anyway. See
growther cache.
Point GROWTHER_CACHE_DIR at a local disk: the cache holds programs C5 loads and models
it reads directly, and a network share that drops out takes them with it. The variable is
read once, when C5 starts, and only from the environment C5 starts in — never from
c5.yaml.
Stop C5 downloading models on its own #
Three features use models that run on your machine, and C5 downloads each one once, the first time it is needed:
| Feature | Download | When C5 downloads it on its own |
|---|---|---|
| Memory search | about 2.3 GB | At start, for any of the three models that has no file yet |
| Offline speech | about 60 MB | At start, while the built-in voice engine is switched on and the model is missing |
| Local classifier | about 812 MB | At start while the classifier is on, and when someone switches it on in Settings › Classifier and saves |
To stop the automatic download of one of them, set its variable to 1:
GROWTHER_QMD_NO_AUTO_DOWNLOAD=1
GROWTHER_VOICE_NO_AUTO_DOWNLOAD=1
GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1Only the value 1 counts. 0, true, yes or an empty value leave the download on.
Spaces around the 1 are ignored.
What each one does and does not do:
- It stops only the download C5 starts on its own. The matching command still
downloads when you run it:
growther qmd-run pull,growther voice pullorgrowther classifier pull. That is the point — the variable is how you say "only when someone asks". - It does not switch the feature off. Memory search's semantic queries, offline speech and the local classifier simply wait until the model is there. You can also place the file by hand; C5 checks it and uses it.
- C5 reads the variable when it would start an automatic download — at start, and for the classifier also when someone switches it on. Restart C5 after changing it.
You can set them in the environment C5 starts in, or as a custom key in c5.yaml (see
Config files).
Two other things stop every automatic download too:
- Setting C5 to Offline (the egress posture, see
Network and egress).
It also refuses the three
pullcommands. - A
CIvariable with any non-empty value, evenfalseor0. C5 never downloads models on its own on a build server. Thepullcommands still work there.
To keep the classifier from downloading anything at all across a fleet, lock the
managed policy key classifier_enabled to false instead. A locked false switches the
feature off and also refuses growther classifier pull. A Recommended (unlocked)
false is only a default: on a machine that has already run a C5 version with the
classifier, the switch is already on and outranks it, and it never refuses pull. See
Managed configuration and
Local classifier.
Local model servers and gateways #
C5 looks for each model server and agent gateway at its usual address. If one runs somewhere else — another port, or another machine — tell C5 where:
| Variable | Service | Default |
|---|---|---|
LMSTUDIO_BASE_URL |
LM Studio | http://127.0.0.1:1234/v1 |
OLLAMA_BASE_URL |
Ollama | http://127.0.0.1:11434/v1 |
LLAMACPP_BASE_URL |
llama.cpp | http://127.0.0.1:8080/v1 |
INFERENCER_BASE_URL |
Inferencer | Hugging Face's cloud, https://api-inference.huggingface.co |
OPENCLAW_GATEWAY_URL |
The OpenClaw gateway | ws://127.0.0.1:18789 |
HERMES_GATEWAY_URL |
The Hermes gateway | http://127.0.0.1:8642 |
HERMES_DASHBOARD_URL |
The Hermes dashboard | Port 9119, on the Hermes gateway's host |
PICOCLAW_GATEWAY_URL |
The PicoClaw gateway | http://127.0.0.1:18790 |
Set them in c5.yaml (see Config files) or in the
environment C5 starts with; the environment wins over c5.yaml. Every connection C5 makes
to that service follows the value in force. A value you remove from c5.yaml can stay in
use until C5 next starts, because C5 keeps what it read when it started.
The model servers. Their addresses are full URLs, such as
http://192.168.1.20:1234/v1. A value C5 cannot read as a URL, or one pointing at a
link-local or cloud-metadata address, is ignored and the default used. Their cards in
Settings › Integrations show the value in force. Pressing Save in Settings stores
the address each model server's card shows, both in c5.yaml under these same names and in
Settings. So once anyone has saved there, removing the variable leaves C5 on the same
address. To go back to the default, clear the address on the card and save (and remove the
variable from the environment, if it is set there).
The gateway variables set a gateway's address and win over the Host and Port on
its card. Give a host and a port only — 10.0.0.4:18789, http://10.0.0.4:18789 or
ws://10.0.0.4:18789 — because C5 builds the rest of the address itself. Always include
the port: without one, C5 uses port 80, not the gateway's usual port.
C5 talks to gateways over plain http and ws, so it ignores a value that starts
https:// or wss:// rather than send the gateway's token unencrypted to an address you
meant to be encrypted. It also ignores a value with a path, a query or a user name and
password in it, a link-local or cloud-metadata address, or a port outside 1–65535. The
card's address, or the default, is then used instead.
To check that C5 took your value, open the gateway's card in Settings ›
Integrations. While a gateway variable is in force, the card shows the address C5 really
uses, the Host and Port it sets cannot be edited, and a note under them names the
variable and says to change or remove it where it is set. If the fields can still be edited
and no note names the variable, C5 is not using your value: either C5 did not start with it,
or it ignored it. When it ignored it, C5's own output says why, once, in a line like
[config] OPENCLAW_GATEWAY_URL is ignored (<why>): <value> — ….
A gateway variable is never copied into Settings, so saving the card does not keep it:
remove the variable, and once C5 restarts the card's own Host and Port apply again,
or the default. HERMES_DASHBOARD_URL moves only the dashboard: without it, the dashboard
is on the Hermes gateway's host — the one HERMES_GATEWAY_URL gives, when that is set.
If you already set OPENCLAW_GATEWAY_URL for OpenClaw. It is also what OpenClaw's own
command-line tool reads to find its gateway, so the two agree when it is set. But C5 started
from a shell that exports it uses it too, and the OpenClaw card's Host and Port lock
(or, for a wss:// address, C5 ignores it as above). This is new after v322: v322 and
earlier set a gateway's address only on its card. If that is not the gateway C5 should use,
start C5 without the variable.
Try early versions #
GROWTHER_CHANNEL=beta growther updateSwitch back with GROWTHER_CHANNEL=stable growther update.
Keep C5 out of your shell files #
Some people manage their PATH by hand and do not want any program editing their shell setup:
GROWTHER_NO_MODIFY_PATH=1 growtherYou will then need to add C5 to your PATH yourself.
Run C5 on your own libvips #
C5 processes images with sharp, which runs on a library called libvips. libvips and the
libraries built into it are licensed under the LGPL, which gives you the right to run C5
with a modified version of them. GROWTHER_LIBVIPS_DIR is how you do that.
1. Lay out a folder the way sharp's own @img folder is laid out, with your build
under these names:
<folder>/sharp-<platform>/lib/sharp-<platform>-<sharp version>.node the sharp binding
<folder>/sharp-libvips-<platform>/lib/libvips-cpp.<version>.dylib macOS
<folder>/sharp-libvips-<platform>/lib/libvips-cpp.so.<version> Linux<platform> is, for example, darwin-arm64, linux-x64 or win32-x64. On Windows
the libvips DLLs, and any other DLLs your libvips needs, go in the same folder as the
binding: <folder>\sharp-win32-x64\lib\.
Instead of that layout, the folder can hold a single sharp-<platform>-<sharp version>.node
that you built from sharp's source against your own libvips.
The exact file names for your copy of C5 — the sharp version and the libvips library
names — are in the libvips section of the third-party notices (growther licenses).
Your libvips must be at least the version that sharp release needs; C5 says so if it is
not.
2. On macOS, sign any library you change (an ad hoc signature is enough) and clear the quarantine flag a download puts on files, or macOS refuses to load them:
codesign --force --sign - <folder>/sharp-libvips-darwin-arm64/lib/libvips-cpp.*.dylib
xattr -dr com.apple.quarantine <folder>3. Start C5 with the variable set:
GROWTHER_LIBVIPS_DIR=/path/to/folder growtherThe first time C5 processes an image it loads your copy and prints a line like this, naming the files the system actually loaded:
[sharp] using libvips <version> from GROWTHER_LIBVIPS_DIR: binding /path/to/folder/sharp-darwin-arm64/lib/sharp-darwin-arm64-<version>.node, library /path/to/folder/sharp-libvips-darwin-arm64/lib/libvips-cpp.<version>.dylibgrowther doctor shows the same thing on its libvips line, when the variable is set
in the terminal you run it from.
If your copy cannot be loaded, C5 does not quietly fall back to its own. It logs the
reason once — for example that the folder is not a directory, that no binding for this
platform is in it, that the binding could not be loaded, or that your libvips is older
than sharp needs — and image features stay off. The message starts
GROWTHER_LIBVIPS_DIR=<folder>: and ends by saying that unsetting the variable restores
the copy bundled with C5. Fix the folder, or unset the variable, and restart C5.
Things to know:
- C5 decides which libvips to use the first time it needs one, and keeps that choice until it restarts. Change the variable, then restart C5.
- The variable is read only from the environment C5 is started in. A
GROWTHER_LIBVIPS_DIRline inc5.yamlor in Settings is refused, because the variable makes C5 load native code. - If C5 starts on its own at sign-in, set the variable where that start reads it — see the next section.
When C5 starts at sign-in #
When you have run growther service install, C5 is started by your operating system, not
by your terminal, so it never sees variables you export in a shell. C5 rewrites its own
start-up definition each time it starts, so put your variables where it leaves them alone:
macOS (launchd). Add the variable to EnvironmentVariables in
~/Library/LaunchAgents/ai.growther.c5.plist. C5 keeps GROWTHER_LIBVIPS_DIR there when
it rewrites that file:
plutil -replace EnvironmentVariables.GROWTHER_LIBVIPS_DIR -string /path/to/folder \
~/Library/LaunchAgents/ai.growther.c5.plistThen sign out and back in, so launchd reads the changed file. Other variables you add there are not kept when C5 rewrites the file.
Linux (systemd). Use a drop-in, which C5 never touches — it only rewrites the main unit:
systemctl --user edit growther-c5Add these lines in the editor that opens, save, then restart C5:
[Service]
Environment=GROWTHER_LIBVIPS_DIR=/path/to/foldersystemctl --user restart growther-c5Without systemd, C5 starts from your desktop's autostart, which uses the environment your desktop session starts in.
Windows. Make it a user environment variable, then sign out and back in:
setx GROWTHER_LIBVIPS_DIR C:\path\to\folderBoth ways C5 starts at sign-in on Windows inherit user environment variables.
The Linux and Windows places work for the other variables on this page too. On macOS C5
keeps only GROWTHER_LIBVIPS_DIR across its rewrite, so for anything else use another
route: the three download switches can go in c5.yaml as custom keys, and the models'
location is better moved with growther cache relocate, which records it for every
start. For ffmpeg, set FFmpeg location in Settings › Tools instead of
FFMPEG_PATH: it is kept in C5's settings, which the rewrite does not touch. Or install
ffmpeg in /opt/homebrew/bin or /usr/local/bin, where C5 looks for it anyway.
What is in your data folder #
Everything C5 knows lives in GROWTHER_HOME (by default ~/.growther):
- Your chats, tasks, and their whole history
- Files your agents made
- Your settings and API keys — see Config files
- Backups
The databases are encrypted, and on macOS and Linux the key file lives in this same folder — so a copy of the folder is a copy of everything, key included. See Local encryption before you back it up anywhere shared.
On Windows it is two folders. Every install created since C5 2026.9 keeps configuration
under %USERPROFILE%\.growther and puts the databases and keys under
%LOCALAPPDATA%\Growther\C5, because %LOCALAPPDATA% does not travel with a roaming
profile — an encrypted database must never be carried to a machine without its key.
The downloaded models #
The cache root — your C5 folder, unless you moved it — holds what C5 downloads and unpacks. None of it is your data, none of it is backed up, and all of it can be fetched again:
| Path under the cache root | What it is |
|---|---|
qmd/cache/qmd/models/ |
The three memory-search models, about 2.3 GB |
qmd/README.md |
Each memory-search model's address, size and SHA-256, for placing them by hand |
speech/ggml-base.en-q5_1.bin |
The offline speech model, about 60 MB |
classifier/models/decider-0.8b.Q8_0.gguf |
The local classifier's model, about 812 MB |
classifier/runtime.json |
What the classifier process reads to find its model. C5 writes it only for a model that passed the check, and writes it again whenever the classifier starts, so it is safe to delete with C5 stopped |
cache/, lib/ |
Other caches, and native libraries C5 unpacks for itself |
Next to a model you may see:
| File | What it means |
|---|---|
<model>.partial |
A download in progress or interrupted. For memory search and the classifier, the next attempt resumes it; the speech model starts again from the beginning |
<model>.partial.lock |
Marks which C5 process is downloading the model right now |
<model>.verified.json |
C5 has checked this file's size and SHA-256, and records the result so it does not read the whole file again |
<model>.rejected (or .rejected.1, …) |
A file that was at the model's name but is not the model. C5 renamed it so it is never loaded, and deleted nothing. Remove it once you no longer need it. (Memory search and offline speech only — a wrong classifier model is left in place, never loaded, and replaced by the next download) |
Run growther cache show to see where these are and how big they are, and
growther home show to see every path your install uses. See
Where your data lives for the full layout.
Locations are decided once #
GROWTHER_HOME and the three folders above are read when C5 starts and never change
while it runs. On a managed device an organisation policy wins over the environment;
otherwise the environment wins over the pointer file C5 writes when you move the home
from Settings. A service unit registered by an older version may still carry an old
GROWTHER_HOME; C5 prefers the pointer in that case and rewrites the unit at the next
start.
Databases are refused on a network folder, a mapped drive, a roaming profile or a synced
folder. GROWTHER_ALLOW_REMOTE_DATA=1 lifts that only in development builds, never in a
released binary.
