Skip to Content
sedge

Configuration

config.json, the files in ~/.sedge, and environment variables.

Most things are set up by asking sedge in the TUI. The rest is ~/.sedge/config.json.

agents, limits and the models are read again whenever they're used, so a change applies at once. The rest (workspace, connections, permissions, sandbox, triageModel, piExtensions, dreaming) is read when the daemon starts: after editing those, restart it with sedge stop && sedge start.

config.json

{
    "workspace": "/Users/you/sedge-work",
    "dreaming": true,
    "permissions": { "bash": "ask", "write": "allow", "mcp": "ask" },
    "limits": { "runsPerHour": 30, "dailyCost": 5, "subagents": 6 },
    "agents": {
        "ci": { "purpose": "Watch the CI signal; tell me only when main breaks." }
    },
    "connections": {
        "telegram": { "kind": "telegram", "settings": {}, "secrets": ["token"] }
    }
}
SettingWhat it doesDefault
modelThe model the daemon's runs use, as provider/model, with :level for thinking. See Models.pi's default
workspaceWhere unwatched runs work (files they create land here).~/.sedge/workspace
sandboxCommands in the OS sandbox: enabled, network, writable, hidden. See Sandbox.off
limitsWhat the agents may use: runs, tokens, money, subagents. See Limits and costs.unlimited
triageModelThe model that triages signals with triage on, as provider/model.see below
dreamingThe nightly memory review. false turns it off. See Memory.on
permissions.bashShell commands in unwatched runs: allow, ask or deny.ask
permissions.writeFile writes and edits in unwatched runs.allow
permissions.mcpMCP tools and other tools in unwatched runs.ask
permissions.signals.bash, .write, .mcpThe same, for runs woken by outside content; the stricter of the two applies. See Security.ask
agentsAgents besides main, by name, each with its purpose. See Agents.none
piExtensionspi extensions to load in the daemon's runs, by name or package. See Extensions.none
connectionsChannels and signals by name. Written by connect; edit with care.none

Each connection has a kind, its settings (never secrets), secrets (the names of the settings kept in the keychain), an optional instruction for a signal, an optional agent that handles it (default main), and enabled: false to pause it. A signal can also have match and ignore (regular expressions an event's text must or must not match), labels and wakeOn (for triage to label each event, and the labels that wake the agent), remember (keep its latest value as a named value instead of waking the agent), when (a condition on its details, like temperature > 30), throttle (how long to stay quiet after an event, like 1h), batch (how long to hold its events and hand them over together, like 5m) and triage: true (a cheap model decides which events need the agent); see Filtering, throttling and triage.

Limits

Every limit is off until you set it. limits is read again whenever config.json changes, so a new limit applies without a restart.

LimitWhat it caps
limits.runsPerHourRuns in the last hour, for each agent with its subagents.
limits.dailyTokensTokens across all agents today, triage included.
limits.dailyCostMoney across all agents today, at the models' listed prices.
limits.subagentsSubagents one agent can have working at once.

What happens when one is reached, and how to spend less to begin with: Limits and costs.

Sandbox

SettingWhat it doesDefault
sandbox.enabledRuns the daemon's commands in the operating system's sandbox.off
sandbox.networkWhether commands may connect anywhere.true
sandbox.writableMore folders it may write to.none
sandbox.hiddenMore files and folders nothing in it may read.none

It's set when the daemon starts; restart after changing it. See Sandbox.

Models

The daemon's runs use pi's default model (the one /model picks in the TUI, at its thinking level) unless you choose another. Three places can, and the most specific wins:

WhereFor
connections.<name>.modelRuns a channel or signal wakes on its own (when it's the only source).
agents.<name>.modelThat agent's runs. See Agents.
modelEvery other run: main, and agents without their own.

Each is provider/model as /model lists it, and may end in a thinking level (off, minimal, low, medium, high, xhigh, max), as --model takes it: "anthropic/claude-opus-5-5:medium". Without one, the run thinks at your default's level. A model you aren't logged in to is skipped, with a warning in the log, and the run uses the default.

You don't need to edit the file: ask the TUI's agent ("run Telegram on Opus at medium"), which uses set_model and asks you first, or send /model on a channel (Channels and signals). They apply from the next run, without a restart. /model in the TUI changes only the TUI's session (and pi's default, which the daemon follows where nothing else is set).

Triage model

Without triageModel, triage uses the cheapest priced model of your default model's provider; if the provider lists no prices (a subscription), one whose name says it's small (mini, haiku, flash, …); and failing that, your default model. Set it ("triageModel": "openai/gpt-5-mini") to choose.

Files

Everything lives in ~/.sedge (or SEDGE_HOME):

PathWhat
config.jsonThe settings above.
sedge.sockThe socket the TUI and the daemon talk over (a named pipe on Windows).
daemon.lockHolds the running daemon's process id, so only one runs.
memory/main's memory, with MEMORY.md as its index; each agent's is in memory/agents/<name>/.
workspace/Where unwatched runs work.
sedge.dbInbox, timers, outbox, approvals.
agent/durable.sqliteThe agent's conversations.
packages/Installed packages of channels and signals.
kinds/Signals and channels you registered.
logs/daemon.log, and the logs and runs sedge dash shows (sedge.jsonl, runs.jsonl).
bin/The standalone install.
secrets.jsonSecrets, only on machines without a keychain.

Model logins, models.json, mcp.json and skills are pi's, in ~/.pi/agent.

Environment variables

VariableWhat it does
SEDGE_HOMEUse another folder than ~/.sedge (a second, separate sedge).
SEDGE_NO_UPDATESet to 1 to only report updates instead of installing them.
SEDGE_AGENT_DIRAnother pi agent folder for the daemon: logins, models, mcp.json, MCP sign-ins, skills.
SEDGE_CODING_AGENT_DIRThe agent folder the TUI uses (logins, models, settings, themes); default ~/.pi/agent.
SEDGE_SECRET_STOREfile keeps secrets in ~/.sedge/secrets.json (owner-only) instead of the OS keychain.
SEDGE_REGISTRYThe npm registry to check for updates.
SEDGE_GITHUB_APIThe GitHub API to check standalone releases against.

On this page