Skip to content

This machine's settings ​

Everything your computer does for Stemma is a setting, kept in one TOML file. The app's This machine page edits it, and so does stemma config. Both keep the file's comments and order.

Scopes ​

Each setting has a general value and can be overridden for one group. In the app, choose Every group or a group at the top of This machine. A group's value says “Set for algebra” and can go back to the general one.

sh
stemma config                                   # every value, and where it comes from
stemma config get agents.claude.model --group algebra
stemma config set agents.claude.model sonnet --group algebra
stemma config unset agents.claude.model --group algebra
stemma config edit                              # opens the file in $EDITOR

The order, highest first, is: command-line options, environment variables, the group's value, the general value, the default.

This machine ​

toml
[machine]
isolation = "docker"    # docker, podman or none
cpus = 4
memory = "8g"
KeyDefaultMeaning
isolation"docker"Where answers and the workshop run: in containers, with docker or podman, or none, on this computer, as you; not recommended
image"stemma/stemma-<version>"The image they run in: Stemma's own, of this version
cpus, memory4, 8gLimits of each container

See the images for where an image comes from.

Agents ​

toml
[agents]
enabled = true          # answer for me
concurrency = 2         # answers at once
priority = "normal"     # high, normal, low, or a number
attachments = "mine"    # none, mine or group

[agents.claude]
model = "opus"

[agents.codex]
args = ["--search"]

[agents.opencode]
model = "openrouter/qwen/qwen3-coder"
timeout_minutes = 30

[groups.seminar.agents]
enabled = false
KeyDefaultMeaning
enabledtrueAnswer for me
concurrency1Answers at once
priority"normal"Which of your computers answers first
attachments"mine"Whose files agents may read
workspaceStemma's data folderWhere each group's folder is

Each of [agents.claude], [agents.codex] and [agents.opencode] takes command, args, model and timeout_minutes, 15 by default.

Workshop ​

toml
[workshop]
enabled = false         # serve as a workshop
build = false           # build the library and papers
build_concurrency = 1
concurrency = 1         # processes at once
runners = []            # all configured, or a subset
others = "ask"          # or "never"

[groups.algebra.workshop]
enabled = true
build = true

The workshop runs where answers do, in the same image and with the same limits, as [machine] says.

Secrets for containers ​

Some agents sign in with an API key, such as OPENROUTER_API_KEY for OpenCode with OpenRouter. Name the variable, and keep its value as a secret:

toml
[agents.opencode.container]
env = ["OPENROUTER_API_KEY"]
sh
stemma config secret set OPENROUTER_API_KEY
stemma config secret list

Secrets are kept in credentials.toml, readable only by you, and passed to the container without appearing on any command line. Variables whose names start with STEMMA_ never enter a container.

Environment variables ​

Every setting has a variable: STEMMA_ and its name in capitals, with dots as underscores, such as STEMMA_AGENTS_CLAUDE_MODEL=opus. Variables apply to every group. STEMMA_HUB and STEMMA_TOKEN sign in without any file.

Folders ​

DefaultHolds
Settings~/.config/stemmaconfig.toml, credentials.toml
State~/.local/state/stemmathe active group, logs, running jobs, approvals
Data~/.local/share/stemmaeach group's folder

STEMMA_CONFIG_DIR, STEMMA_STATE_DIR and STEMMA_DATA_DIR move them.

The group's folder ​

text
library/         the library at main
papers/<slug>/   each paper at main

Stemma manages this folder and brings it up to date before answers and --workspace sessions. Look at it, but do not keep your own work there. Lean's build files are kept between runs, so Mathlib is not built again.

Stemma is early: it changes often, and breaks.