Config

subs.toml declares one project. Every key is here.

The CLI reads the file from the working directory, or from the path you pass with -c. It does not search parent directories.

An unknown key is a parse error.

A full file

name = "support-bot"
db = "subs.db"
log = "info"

[llm.claude]
type = "anthropic"

[agent.support]
llm = "claude"
model = "claude-sonnet-4-5"
system = "You are a support agent."
mcp = ["mcp.sentry"]
subagents = ["agent.researcher"]

[agent.researcher]
description = "Finds and reads sources."
llm = "claude"
model = "claude-haiku-4-5"

[mcp.sentry]
url = "https://mcp.sentry.dev/mcp"

[agent.support.slack]
name = "Support"

[serve]
host = "127.0.0.1"
port = 8080

[remote]
url = "https://api.substructure.ai"
org = "org_01hx…"
project = "proj_01hx…"

What a file describes

A file has two roles. It can have one or both.

The project itself stays the same for both roles: name, [llm.<id>], [agent.<id>], [mcp.<id>], [plugin.<id>], and [skill.<id>].

A second environment is a second file. subs apply -c subs.staging.toml deploys a separate project.

Top level

Unset, db is ~/.config/subs/subs.db, beside credentials.toml. Every command on the machine reads that one, whichever directory it runs in and whether or not a file is there. Set db to give a project its own. Two files that both set it are two engines.

[llm.<id>]

Where a model call runs. An agent names a block by its ID.

[llm.claude]
type = "anthropic"
api_key_env = "MY_ANTHROPIC_KEY"

[llm.byo]
type = "worker"
format = "anthropic"

There is no default block and no fallback. An agent names a block, or its calls fail. See LLMs.

A worker block takes no api_key_env, base_url, or cache_ttl. The call never leaves your worker.

api_key_env names a variable on your machine. subs apply removes it. A deployment refuses a document that carries one.

[system1.<id>]

Where a system1.call runs. A worker names a block by its ID.

[system1.cf]
type = "workers-ai"
model = "clef-flash"
account_id = "0123abcd…"
price = { input = "0.038" }

See System One models.

[agent.<id>]

An agent. The ID is what clients, channels, and parent agents route on.

[agent.support]
llm = "claude"
model = "claude-sonnet-4-5"
system = "You are a support agent."
mcp = ["mcp.sentry"]
subagents = ["agent.researcher"]
tools = [{ name = "confirm", description = "Ask a person", handler = "client" }]
worker = "https://bot.example.com/agent"
signing_secret_env = "SUPPORT_SIGNING_SECRET"

[agent.support.retry]
tool = { max_attempts = 3 }

An agent that sets no config needs a worker. An agent that sets any config needs llm and model. See Agents.

The tools your worker runs can go in the file, but they do not have to: the worker can return them in the config. An agent with a tool, a skill, or a plugin that the worker answers needs a worker.

mcp and plugins entries

Each entry is an ID on its own, or a table. mcp takes a connection path such as mcp.sentry or plugin.pdf.mcp.renderer. plugins takes a plugin ID.

[agent.support]
mcp = [
  "mcp.sentry",
  { id = "mcp.linear", tools = { read_only = true }, approve = "destructive" },
]

On a plugins entry these settings apply to each of the plugin's MCP servers.

Defer every tool

defer_tools = true takes the defaults. A table sets them. The presence of the key is the switch, so an agent cannot carry a setting that does nothing. A tool or a connection overrides it with its own defer.

See Deferred tools.

[agent.<id>.attachments]

How each file the agent receives reaches the model. Without this section every file rides in the prompt as media, and the agent offers no attachment tool.

[agent.analyst.attachments]
tools = ["read", "view"]
max_inline = "20mb"

[agent.analyst.attachments.rules]
"image/*"         = "inline"
"application/pdf" = "inline"
"*/*"             = "attachment"

A key in rules is a mime pattern: an exact type (application/pdf), a type wildcard (image/*), or */*. The most specific match wins. A file no pattern matches rides as media, the way it does without this section.

inline keeps the file in the prompt and adds one line naming it. attachment keeps the line alone. The line reads [attachment sales.csv text/csv 2.1 MB]. The id is the file name, with a counter when the name is already taken on this conversation. The bytes are stored in both cases, and the model reads them with attachment_read (text) or attachment_view (image, audio, and video).

The decision is made once, when the file arrives, and never changes.

The same table is attachments on the agent config, so a worker can set it at session.start like any other field. The rules are applied when a message is delivered to the worker, after session.start has settled the config, so they cover the first message of a session too.

[agent.<id>.context]

How the agent makes room in what the model reads. Before a model call, the engine tries the strategies in order. See Make room.

[agent.support.context]
window = 200000

[[agent.support.context.strategy]]
use = "reset"
idle_secs = 3600

[[agent.support.context.strategy]]
use = "summarize"
keep_tokens = 20000

Each strategy:

The same table is context on the agent config.

[events.<id>]

A URL that gets the events and the live text of the sessions of each agent that names it. See Events URL.

[events.app]
url = "https://app.example.com/events"

[mcp.<id>]

An MCP server the engine connects to.

[mcp.sentry]
url = "https://mcp.sentry.dev/mcp"

[mcp.github]
url = "https://api.githubcopilot.com/mcp/"
auth = "token"
prefix_tools = false

A token written in the file is a parse error. Fill a connection with subs auth <path>. See Connectors.

[plugin.<id>]

A plugin directory the CLI resolves and sends to the deployment, or a plugin whose skills the worker answers. An agent names a plugin by its ID. An ID that starts with _ is the engine's.

[plugin.pdf]
path = "./plugins/pdf-tools"

[plugin.pdf.mcp.renderer]
auth = "none"

The CLI resolves the directory to data, at startup for a local engine and at subs apply for a deployment, so a session never reads plugin files. A plugin's servers join the connection registry as plugin.<id>.mcp.<server> and authorize like any connection. The model sees their tools under <plugin>_<server>__<tool>. See Plugins.

[skill.<id>]

A skill of the project's own: a directory with a SKILL.md, or a skill that the worker answers. An agent names it in skills. See Skills.

[skill.automations]
path = "./skills/automations"

[skill.visits]
source = "worker"
description = "Who visited the clinic, and when"
tools = ["find_bookings"]

[plugin.<id>.mcp.<server>]

What this deployment says about one server the plugin declares, keyed by its name in the plugin's mcp.json. Every key overrides the plugin's. One left out keeps what the plugin shipped.

[plugin.pdf.mcp.renderer]
auth = "token"
url = "https://pdf.staging.example.com/mcp"

Authorize it by its path: subs auth plugin.pdf.mcp.renderer.

[agent.<id>.slack]

Gives one agent its own Slack app. The block being there is the whole declaration.

[agent.support.slack]
name = "Support"
description = "Answers customer questions"
answers = "both"

Every agent with this block is a separate Slack app. Which channels it answers in is decided in Slack, by who invites it.

Run subs auth agent.<id>.slack to set it up. See Slack.

[serve]

Defaults for subs serve.

max_body covers what reaches the engine: a client message and a worker's decision answer. It does not cover what the engine sends a worker. A worker that hosts the model receives the media in the prompt, so its own server must accept a body that size.

[remote]

The deployment this file administers. That can be the hosted cloud, one you host, or another person's subs serve.

subs apply writes the pin back into the file and keeps your comments. A second apply changes nothing.

Precedence

flag > environment variable > subs.toml > default

Setting a value in the file still lets you override it on the command line.

The message, --input, and --session have no key in the file. They say what one run does.

Secrets

The file names secrets. It never holds them.

subs apply strips api_key_env and signing_secret_env before it sends.

Next steps

• Agents: what an agent section declares.
• Plugins: what a [plugin.<id>] directory holds.
• Skills: what a [skill.<id>] section declares.
• CLI: the commands that read this file.
• Cloud: applying it to a deployment.
• Protocol: the same types on the wire.