Skip to content

Agent Profiles

Profiles bundle model choice, prompt behavior, tool access, read-only posture, turn limits, and reasoning effort. Pick the profile that matches the job.

ProfilePromptKindToolsReadOnlyMaxTurnsDefaultEffortUse Case
craftchatgeneral coding and extension toolsfalse203General coding, features, bug fixes
quickquickfocused coding/search subsetfalse102Small, fast, well-defined tasks
studystudyextended reasoning and coordination toolsfalse404Architecture, debugging, multi-step changes
reviewreviewread-only review tools plus diagnosticstrue123Code review (read-only)
oracleoracleread-only investigation toolstrue84Advisory second opinion (read-only)
findfindread, multi_read, finder, glob, grep, web_search, todo_readtrue82Repository and manual/project-note discovery
aggmanaggmanconfigured: find_thread, read_thread, web_search, send_message, receive_messages, todo_read, todo_write, ask_user
effective: find_thread, read_thread, web_search, receive_messages, todo_read
true62Multi-agent aggregation and status synthesis

Aliases:

AliasCanonical
smartcraft
rushquick
deepstudy
searchfind

Each profile uses a promptKind that shapes the agent’s behavior:

  • chat - General-purpose assistant. Balances reasoning with action.
  • quick - Optimizes for small, fast tasks. Fewer tools, lower turn budget.
  • study - Uses deeper reasoning for architecture, debugging, and multi-step changes.
  • review - Reviews code and plans critically. Read-only posture.
  • oracle - Acts as an advisory second brain. Read-only, high reasoning effort.
  • find - Focuses on repository and manual/project-note discovery. Read-only.
  • aggman - Receives and synthesizes coordination messages in multi-agent setups. Because the built-in profile is read-only, configured state-writing and interactive tools such as send_message, todo_write, and ask_user are filtered out of its effective registry.
  • /config - open the profile and model picker.
  • /mode - cycle through craft -> quick -> study.
anode -x --profile study "debug this crash"
anode execute --profile quick "add input validation to signup handler"
anode review --profile review
anode tools list --profile review

The --profile flag is wired into root -x, anode execute, daemon runs start, anode review, and anode tools inspection/use commands.

Define custom profiles in your config.json under profiles:

{
"profiles": {
"security-review": {
"model": "anthropic/claude-sonnet-4-6",
"promptKind": "review",
"systemPrompt": "You are a security-focused code reviewer. Flag injection vulnerabilities, auth bypasses, data exposure, and unsafe deserialization. Cite CWE IDs where applicable.",
"tools": ["read", "finder", "web_search", "todo_read"],
"readOnly": true,
"maxTurns": 15,
"defaultEffort": 4
}
}
}

Use it:

anode review --profile security-review
FieldTypeDescription
namestringProfile name
modelstringModel override (e.g., "anthropic/claude-sonnet-4-6")
promptKindstringPrompt kind: chat, quick, study, review, oracle, find, or aggman. Aliases: rush -> quick, deep -> study, search -> find.
promptPathstringPath to a custom system prompt file
systemPromptstringInline custom system prompt text
toolsstring[]Allowed tools. Supports glob patterns. ["*"] means all
readOnlyboolRead-only posture. Omits mutation tools from the model’s view
maxTurnsintMaximum model/tool turns per run
defaultEffortintDefault reasoning effort (1-5)

Custom profiles merge over built-in defaults. If you define a profile named quick, your fields override the built-in quick values. Missing fields keep their defaults.

Profile names are case-insensitive. When you define a new custom profile name, Anode uses craft as its merge base. When you request a profile at runtime, that name must exist as either a built-in, configured custom profile, or alias. Alias names are resolved before lookup.

Profiles with readOnly: true omit mutation tools from the model-visible registry. The agent cannot see or call edit_file, create_file, or bash with write operations. This is enforced at the tool registry level - the model never receives these tools in its schema.

The built-in review, oracle, find, and aggman profiles are read-only.

Each model registers capabilities that interact with profiles:

CapabilityDescription
SupportsReasoningModel supports extended reasoning/thinking
DefaultEffortDefault reasoning effort level (1-5)
MinEffortMinimum supported effort
MaxEffortMaximum supported effort
ThinkingRequiredModel requires thinking to be enabled

The profile’s defaultEffort is clamped to the model’s MinEffort-MaxEffort range. Cycle effort in the TUI with /effort.