Tools

Tools are how Anode touches the world: read files, search code, run commands, edit files, call MCP servers, spawn subagents, and search the web. You control which tools exist and what they can do with permissions and configuration.

External fetch is a thin surface: web_search is the only first-class network primitive, and bash + curl is the recommended fallback for raw page bodies. The OpenAI shell_command request shape is accepted via canonical-name aliasing and dispatched to bash.

How Tool Calls Work

1. The model emits a TOOL: directive with a JSON payload containing the tool name and arguments.
2. Anode validates the call against your permission policy.
3. If approved, Anode executes the tool and captures the result.
4. The result returns to the model as context for the next turn.

TOOL: {"name": "read", "args": {"path": "src/main.go"}}

Tools require different permission levels. Read-only tools run automatically. Write tools and shell commands require confirmation unless you set a permissive approval mode.

Permission Levels

Level	Behavior
auto_read	Always allowed. No confirmation needed.
external_read	External service reads and read-only delegated investigation.
state_write	Local agent state mutations.
confirm_execute	Requires confirmation for shell execution.
confirm_write	Requires confirmation for file mutations.
delegated	Delegated to subagent or skill.
interactive	Requires direct user interaction.

Built-In Tools

The unfiltered built-in registry loads the core tool set before dynamic language-server tools. Profiles then filter that registry by allowlist and read-only policy: craft exposes the normal working set, review is read-only, and study adds coordination tools. Autoresearch tools are present for TUI research workflows. The lsp and get_diagnostics tools register only when a language server is configured or auto-detected. MCP, toolbox, skill-local MCP, and process-plugin tools are registered dynamically.

bash

Run a shell command in the workspace.

Permission: confirm_execute

Parameter	Type	Required	Default	Description
command	string	yes	-	Shell command to execute
cwd	string	no	workspace root	Working directory
timeout_ms	int	no	30000	Command timeout in milliseconds. Values above 1800000 are reduced to 1800000.

TOOL: {"name": "bash", "args": {"command": "go test ./...", "timeout_ms": 60000}}

read

Read a text file with line numbers. By default, returns the first 1000 lines. Use read_range, start_line, and end_line to slice large files.

Permission: auto_read

Parameter	Type	Required	Default	Description
path	string	yes	-	File path relative to workspace
read_range	array	no	-	1-indexed positive [start, end] range. Length 1 reads from start through the end of file. Takes precedence over start_line and end_line.
start_line	int	no	1	First line to read
end_line	int	no	end of file	Last line to read
max_bytes	int	no	2097152	Byte ceiling on numbered output after line slicing.

read truncates in two stages: it first reads up to a 2 MiB file prefix, then applies max_bytes to the numbered slice. The structured result sets file_capped for the prefix limit and truncated if either stage trims output. Use read_range, start_line, or end_line for normal large-file focusing; max_bytes is mainly a guard against single huge lines.

TOOL: {"name": "read", "args": {"path": "internal/app/app.go", "start_line": 1, "end_line": 50}}

multi_read

Read up to 10 text files in one round-trip. Each file is rendered as if by read (numbered lines, same 2 MiB per-file cap) and separated by ===== <path> ===== banners. Use this instead of emitting N parallel read calls when you already know the paths.

Permission: auto_read

Parameter	Type	Required	Default	Description
paths	array	yes	-	Workspace-relative file paths (max 10).
read_range	array	no	-	1-indexed [start, end] slice applied to every path.
start_line	int	no	-	Alternative to read_range; applied to every path.
end_line	int	no	-	Alternative to read_range; applied to every path.
max_bytes	int	no	4194304	Total byte ceiling across all files. Files past the ceiling are skipped with a truncation note.

A failure on one path is recorded inline (banner plus [error: ...]) and the other paths still return; the structured result lists per-file metadata under files.

TOOL: {"name": "multi_read", "args": {"paths": ["go.mod", "go.sum"]}}
TOOL: {"name": "multi_read", "args": {"paths": ["internal/tools/read.go", "internal/tools/multi_read.go"], "read_range": [1, 40]}}

edit_file

Edit an existing text file. Provide exactly one input shape: edits, old_str/new_str, or unified_diff.

Permission: confirm_write

Parameter	Type	Required	Default	Description
path	string	yes	-	File path relative to workspace
edits	array	one of edits/old_str/unified_diff	-	Array of edit operations
old_str	string	one of edits/old_str/unified_diff	-	Text to replace in top-level single-replacement form
new_str	string	with old_str	-	Replacement text for old_str
replace_all	bool	no	false	Replace every old_str occurrence; otherwise old_str must be unique
unified_diff	string	one of edits/old_str/unified_diff	-	Unified diff patch to apply

Each item in edits:

Field	Type	Required	Description
old	string	yes	Text to find
new	string	yes	Replacement text
occurrence	int	no	Which occurrence to replace, 1-indexed. 0 means the old text must be unique.

TOOL: {"name": "edit_file", "args": {"path": "main.go", "edits": [{"old": "fmt.Println(\"hello\")", "new": "fmt.Println(\"goodbye\")", "occurrence": 1}]}}

apply_patch

Apply a unified diff to the workspace. Provide the diff in patch; input and diff are accepted as aliases. Anode validates the patch before approval.

Permission: confirm_write

Parameter	Type	Required	Default	Description
patch	string	yes	-	Unified diff. input and diff are accepted as aliases for compatibility.

TOOL: {"name": "apply_patch", "args": {"patch": "--- a/main.go\n+++ b/main.go\n@@ -1 +1 @@\n-old\n+new\n"}}

create_file

Create a new text file. Fails if the file exists unless overwrite is true.

Permission: confirm_write

Parameter	Type	Required	Default	Description
path	string	yes	-	File path relative to workspace
content	string	yes	-	File content
overwrite	bool	no	false	Overwrite existing file

TOOL: {"name": "create_file", "args": {"path": "scripts/deploy.sh", "content": "#!/bin/bash\nset -e\ngo build -o bin/app ./cmd/app"}}

finder

Find files, text, symbols, or behavioral patterns across the workspace.

Permission: auto_read

Parameter	Type	Required	Default	Description
mode	string	no	”text”	Search mode: files, text, symbols, or agent
query	string	no	-	Search query or pattern
path	string	no	workspace root	Directory to search
glob	string	no	-	File glob filter
limit	int	no	50	Returned result limit.
case_sensitive	bool	no	false	Case-sensitive matching

mode=agent delegates behavioral or cross-module searches to a read-only search subagent. If delegation is unavailable or fails, Anode falls back to text search and sets agent_fallback in the structured result data.

TOOL: {"name": "finder", "args": {"mode": "text", "query": "func NewServer", "glob": "*.go"}}

glob

Match files by glob pattern.

Permission: auto_read

Parameter	Type	Required	Default	Description
pattern	string	yes	-	Glob pattern such as **/*.go. Supports recursive **.
path	string	no	.	Workspace-relative root.
limit	int	no	50	Returned match limit.
include_hidden	bool	no	false	Include dot-prefixed entries.

TOOL: {"name": "glob", "args": {"pattern": "**/*_test.go", "path": "internal", "limit": 100}}

grep

Search file contents with ripgrep when available, with a POSIX grep fallback. Captured stdout is limited to 256 KiB and marked truncated when exceeded.

Permission: auto_read

Parameter	Type	Required	Default	Description
pattern	string	yes	-	Literal string unless regex is true.
path	string	no	.	Workspace-relative directory.
glob	string	no	-	File glob filter, such as *.go.
case_sensitive	bool	no	false	Match case exactly.
word_match	bool	no	false	Whole-word match.
regex	bool	no	false	Treat pattern as regex.
max_matches	int	no	50	Returned match limit per file.
context	int	no	0	Context lines around each match.

TOOL: {"name": "grep", "args": {"pattern": "TODO", "glob": "*.go", "context": 1}}

render

Emit a structured render spec the TUI displays as rich components (cards, tables, metrics, status lines, etc.). Headless consumers receive the same JSON in tool result data.render_spec so the output round-trips outside the TUI.

Permission: auto_read

Parameter	Type	Required	Default	Description
root	string	yes	-	ID of the root element in elements.
elements	object	yes	-	Map of element id → {type, props, children}. children is an array of element id strings. Types are PascalCase component names.

Supported component types: Box, Text, Heading, Divider, Newline, Spacer, BarChart, Sparkline, Table, List, Card, StatusLine, KeyValue, Badge, ProgressBar, Metric, Callout, Timeline.

TOOL: {"name": "render", "args": {"root": "card", "elements": {"card": {"type": "Card", "props": {"title": "Health"}, "children": ["s"]}, "s": {"type": "StatusLine", "props": {"text": "API", "status": "success"}, "children": []}}}}

chart

Render a small terminal chart from structured data.

Permission: auto_read

Parameter	Type	Required	Default	Description
type	string	no	bar	bar, line, sparkline, or table.
title	string	no	-	Optional title.
data	object	one of data/series	-	Map of label to value.
series	array	one of data/series	-	Ordered array of {label,value} points.
width	int	no	40	Chart width for bar/line charts.

TOOL: {"name": "chart", "args": {"type": "bar", "title": "latency", "data": {"p50": 12, "p95": 45}}}

task

Spawn a bounded subagent for independent work. The subagent runs in its own context with limited tools and turns. Child tasks cannot spawn further delegated agents; task, oracle, and code_review are filtered even when requested in tools_allowed.

Permission: delegated

Parameter	Type	Required	Default	Description
title	string	no	-	Task title for display
goal	string	yes	-	What the subagent should accomplish
context	string	no	-	Additional context for the subagent
tools_allowed	[]string	no	profile default	Tool names or globs the subagent may use; recursive delegation tools are filtered
max_steps	int	no	20	Maximum turns. Values above 40 are reduced to 40.
max_tool_calls	int	no	max_steps * 3	Total tool-call limit
profile	string	no	”craft”	craft default, quick fast/shallow, study thorough/deep, find retrieval-focused, review critique, or oracle read-only advisor. Aliases: smart -> craft, rush -> quick, deep -> study, search -> find.
read_only	bool	no	false	Restrict to read-only tools
timeout_seconds	int	no	300	Wall-clock timeout in seconds. Values above 1800 are reduced to 1800.

TOOL: {"name": "task", "args": {"title": "Audit auth middleware", "goal": "Review all authentication middleware for security issues", "read_only": true, "max_steps": 30}}

oracle

Consult a second-opinion model for complex reasoning. Useful for architecture decisions or debugging.

Permission: external_read

Parameter	Type	Required	Default	Description
question	string	yes	-	Question to ask the oracle model
context	string	no	-	Additional context
files	[]string	no	-	File paths to include as context
focus	string	no	-	Area to focus reasoning on
model	string	no	-	Specific model to use
effort	string	no	medium	none, minimal, low, medium, high, xhigh, or max
timeout_seconds	int	no	180	Oracle runtime timeout in seconds. Values above 600 are reduced to 600.

TOOL: {"name": "oracle", "args": {"question": "Is this concurrency pattern safe?", "files": ["internal/run/engine.go"], "focus": "race conditions"}}

code_review

Run a structured read-only review pass over a described diff or file set. The reviewer uses read/search tools, not shell diff commands, so include files when exact file context matters.

Permission: external_read

Parameter	Type	Required	Default	Description
diff_description	string	yes	-	Natural-language description of the diff or change, with enough context for a read-only reviewer.
files	[]string	no	all relevant	Specific files to focus.
instructions	string	no	-	Extra reviewer guidance.
check_scope	string	no	-	Check scope such as all or changes-only.
check_filter	[]string	no	all	Check names to run.
checks_only	bool	no	false	Return only structured check results.
thinking	string	no	low	low or high.

TOOL: {"name": "code_review", "args": {"diff_description": "the uncommitted changes in internal/run", "files": ["internal/run/engine.go"], "thinking": "high"}}

web_search

Search the web for current information.

Permission: external_read

Parameter	Type	Required	Default	Description
query	string	yes	-	Search query
max_results	int	no	10	Results to return. Values above 20 are reduced to 20.
recency	string	no	any	any, day, week, month, or year. Shorthand such as 24h, 1d, 7d, 30d, and 365d is accepted and reported back as the canonical window.
domains	[]string	no	-	Restrict to specific domains
exclude_domains	[]string	no	-	Exclude specific domains
category	string	no	-	Exa hint: company, people, news, research, or blog
start_published_date	string	no	-	Exa-only lower publish-date bound (YYYY-MM-DD)
end_published_date	string	no	-	Exa-only upper publish-date bound (YYYY-MM-DD)

TOOL: {"name": "web_search", "args": {"query": "Go 1.23 iterator proposal", "max_results": 3, "domains": ["go.dev"]}}

find_thread

Search local Anode threads and sessions.

Permission: auto_read

Parameter	Type	Required	Default	Description
query	string	yes	-	Search query
source	string	no	”local”	Thread source
limit	int	no	10	Results to return. Values above 50 are reduced to 50.

TOOL: {"name": "find_thread", "args": {"query": "refactor database layer"}}

read_thread

Read content from a specific thread.

Permission: auto_read

Parameter	Type	Required	Default	Description
source	string	no	”local”	Thread source
thread_id	string	yes	-	Thread identifier
goal	string	no	-	Reading goal (helps summarize)
max_messages	int	no	40	Messages to return. Values above 80 are reduced to 80.

TOOL: {"name": "read_thread", "args": {"thread_id": "sess_abc123", "goal": "find the auth refactor decision"}}

handoff

Create a structured handoff package for another agent or the user.

Permission: delegated

Parameter	Type	Required	Default	Description
target	string	no	”user”	Handoff recipient: user, next_thread, parent, or aggman
reason	string	no	-	Why the handoff is needed
summary	string	yes	-	Summary of work completed
goal	string	no	-	Single-sentence next-task statement
completed	[]string	no	-	List of completed items
open_questions	[]string	no	-	Unresolved questions
recommended_next_steps	[]string	no	-	Suggested follow-up actions
artifacts	[]string	no	-	Relevant file paths or references
relevant_files	[]string	no	-	Files the receiver should open before starting
follow	bool	no	false	Navigate to the spawned thread when supported

TOOL: {"name": "handoff", "args": {"summary": "Migrated auth to JWT tokens", "completed": ["Updated middleware", "Added token refresh"], "open_questions": ["Should refresh tokens expire?"]}}

look_at

Inspect visual or UI artifacts. Provide either path or target.

Permission: auto_read

Parameter	Type	Required	Default	Description
target	string	one of path/target	-	Named visual target
path	string	one of path/target	-	Path to image or visual file
question	string	no	-	What to look for

TOOL: {"name": "look_at", "args": {"path": "manual/screenshot.png", "question": "Is the sidebar alignment correct?"}}

list_skills

List available built-in, user, and workspace skills so the model can choose a valid skill name before loading instructions.

Permission: auto_read

No parameters.

TOOL: {"name": "list_skills", "args": {}}

skill

Load a predefined Anode skill. The tool returns the skill’s instructions to the model and can activate skill-local MCP tools. It does not create a child run by itself. See AGENTS.md and Skills for skill configuration.

Permission: delegated

Parameter	Type	Required	Default	Description
name	string	yes	-	Skill name
input	string	no	-	Input for the skill

TOOL: {"name": "skill", "args": {"name": "code_review", "input": "Review the changes in internal/auth/"}}

ask_user

Pause and ask the user a structured question.

Permission: interactive

Parameter	Type	Required	Default	Description
question	string	yes	-	Question text
reason	string	no	-	Why this information is needed
options	[]object	yes	-	Array of answer options. Each option requires id and label; description is optional.
allow_freeform	bool	no	false	Allow typed response beyond options
required	bool	no	false	User must answer before continuing
topic	string	no	-	Optional navigation/topic label. Multi-word topics are hyphenated.
questions	[]object	no	-	Optional follow-up questions asked together as a questionnaire (max 3 follow-ups; root question + up to 3). Each entry has question, options, and optional reason/topic/allow_freeform/required.

TOOL: {"name": "ask_user", "args": {"question": "Which database driver?", "options": [{"id": "postgres", "label": "PostgreSQL"}, {"id": "sqlite", "label": "SQLite"}], "allow_freeform": true}}

todo_read

Read the current agent task list.

Permission: auto_read

No parameters.

TOOL: {"name": "todo_read", "args": {}}

todo_write

Replace the current agent task list.

Permission: state_write

Parameter	Type	Required	Default	Description
todos	[]object	yes	-	Array of task items

Each item in todos:

Field	Type	Required	Description
id	string	no	Unique task identifier. Auto-filled when empty.
content	string	yes	Task description
status	string	no	Status: pending, in_progress, or completed. Aliases such as todo, active, and done normalize to these values.
active_form	string	no	Optional short phrase for the in-progress form. activeForm is accepted as an input alias.

TOOL: {"name": "todo_write", "args": {"todos": [{"id": "1", "content": "Fix auth bug", "status": "in_progress"}, {"id": "2", "content": "Add tests", "status": "pending"}]}}

task_list

Persist fine-grained tasks with stable IDs, dependencies, parent/child structure, and status.

Permission: state_write

Parameter	Type	Required	Default	Description
action	string	yes	-	create, list, get, update, or delete.
taskID	string	for get/update/delete	auto on create	Stable task ID.
title	string	for create	-	Short title.
description	string	no	-	Longer detail or acceptance criteria.
repoURL	string	no	current workspace	Repository URL/filter.
status	string	no	open on create	open, in_progress, or completed.
dependsOn	[]string	no	-	Blocking task IDs.
parentID	string	no	-	Parent task ID.
limit	int	no	0	Returned task limit for list; 0 returns all matches.

TOOL: {"name": "task_list", "args": {"action": "create", "title": "Document provider retries"}}

task_list update treats omitted optional fields as “leave unchanged” and explicit blank description, repoURL, or parentID values as clear operations. dependsOn: [] clears dependencies. The task graph rejects new parent/dependency cycles, including direct self-references.

lsp

Code intelligence via language servers. Provides go-to-definition, references, hover, and more.

Permission: auto_read

Availability: conditional. The tool appears only when Anode finds a configured or auto-detected language server. See Language Server Protocol.

Parameter	Type	Required	Default	Description
operation	string	yes	-	LSP operation to perform
path	string	no	-	File path (required for most operations)
line	int	no	-	0-indexed line number
character	int	no	-	0-indexed column number
query	string	no	-	Search query (for symbol operations)
includeDeclaration	bool	no	false	Include declaration in references
limit	int	no	200	Returned result limit. Values above 1000 are reduced to 1000.

Operations:

Operation	Description
goToDefinition	Jump to symbol definition
findReferences	Find all references to symbol
hover	Get type or symbol hover info
documentSymbol	List symbols in a file
workspaceSymbol	Search symbols across workspace
goToImplementation	Find interface implementations
diagnostics	Get file diagnostics
prepareCallHierarchy	Prepare call hierarchy item
incomingCalls	Find callers of a function
outgoingCalls	Find callees of a function

TOOL: {"name": "lsp", "args": {"operation": "findReferences", "path": "internal/run/engine.go", "line": 42, "character": 10}}

get_diagnostics

Read diagnostics from configured or auto-detected language servers. This registers with lsp.

Permission: auto_read

Parameter	Type	Required	Default	Description
path	string	no	-	Single workspace-relative file path.
paths	[]string	no	-	Multiple workspace-relative file paths. Combined with path.
wait	bool	no	false	Wait for fresh diagnostics.
wait_timeout_ms	int	no	10000	Wait timeout in milliseconds. Values above 30000 are reduced to 30000.
severity	string	no	any	any, error, warning, info, or hint.

TOOL: {"name": "get_diagnostics", "args": {"path": "internal/tools/web_search.go", "severity": "error"}}

send_message

Post an in-process coordination message to a named mailbox. study can send and receive messages; the built-in aggman profile is read-only, so it can receive and summarize messages but cannot send them.

Permission: state_write

Parameter	Type	Required	Default	Description
recipient	string	yes	-	Mailbox name such as aggman, parent, broadcast, or a worker label. JSON calls may use to as an alias.
message	string	yes	-	Message body.
type	string	no	status	status, question, result, error, or broadcast.
from	string	no	anode	Sender label.
metadata	object	no	-	Free-form key/value metadata.

TOOL: {"name": "send_message", "args": {"recipient": "aggman", "message": "phase 1 complete", "type": "result", "from": "worker-a"}}

receive_messages

Drain or peek at coordination messages addressed to a named recipient. Use this as an aggregator mailbox reader, or as a worker checking for instructions from a parent process.

Permission: auto_read

Parameter	Type	Required	Default	Description
recipient	string	yes	-	Mailbox name to read.
peek	bool	no	false	Return messages without removing them.

TOOL: {"name": "receive_messages", "args": {"recipient": "aggman"}}

init_experiment

Initialize an autoresearch session. The session records a benchmark command, the primary metric, and whether higher or lower values are better. The normal loop is init_experiment, then repeated run_experiment and log_experiment calls.

Permission: confirm_write

Parameter	Type	Required	Default	Description
name	string	yes	-	Short label such as exa-cache-tuning.
objective	string	no	-	Plain-English statement of what should improve.
primary_metric	string	yes	-	Metric name such as latency_ms or hit_rate.
unit	string	no	-	Display unit such as ms or %.
direction	string	yes	-	lower or higher.
command	string	no	-	Benchmark command that emits the metric.
scope	array	no	-	Informational files or directories in scope.
max_runs	integer	no	-	Experiment run limit before the session stops; 0 or omitted means no run-count limit.
reset	boolean	no	false	Start a fresh session even when one is active.

TOOL: {"name": "init_experiment", "args": {"name": "exa-tuning", "primary_metric": "p95_ms", "direction": "lower", "command": "go test -bench=BenchmarkExaSearch ./internal/tools/...", "max_runs": 5}}

run_experiment

Run the configured benchmark command, parse the primary metric, and create a pending result. Every run_experiment call must be followed by log_experiment to commit the result.

Permission: state_write

Parameter	Type	Required	Default	Description
command	string	no	configured command	Override command for this run.
timeout_ms	integer	no	configured timeout	Override per-run timeout.
checks_timeout_ms	integer	no	configured checks timeout	Timeout for post-run checks.

TOOL: {"name": "run_experiment", "args": {}}

log_experiment

Commit the pending autoresearch run. Use baseline for the first reference run, keep for an improvement to retain, discard for a change to drop, crash for a failed benchmark, and checks_failed when validation failed.

Permission: state_write

Parameter	Type	Required	Default	Description
status	string	yes	-	baseline, keep, discard, crash, or checks_failed.
metric_value	number	no	parsed metric	Override the parsed metric value.
description	string	no	-	What the run changed or measured.
commit_message	string	no	-	Git commit subject when status=keep.

TOOL: {"name": "log_experiment", "args": {"status": "baseline", "description": "reference timing"}}

Inspect The Actual Registry

The authoritative list is the current registry:

anode tools list
anode tools show bash
anode tools list --profile review

Use --profile to see what a specific profile can actually call after allowlists, read-only filtering, disabled tools, web access settings, MCP, toolbox, and plugin discovery are applied.

Invoke Read-Only Tools From The CLI

tools use runs read-only tools directly (auto_read and external_read):

anode tools use read --arg path=README.md
anode tools use read --json '{"path":"README.md","max_bytes":4000}'
anode tools use chart --arg type=bar --arg-json data='{"fixed":3,"open":1}'

Use --allow-non-read only when you intentionally want direct invocation of state-writing, file-writing, execute, delegated, or interactive tools.

Dynamic Tools

Beyond built-in tools, Anode supports dynamically registered tools from external systems and skills.

MCP tools

Tools discovered from MCP servers. Named mcp__<server>__<tool>.

TOOL: {"name": "mcp__github__create_issue", "args": {"title": "Bug report", "body": "..."}}

Toolbox tools

Executable scripts in toolbox directories. Named tb__<name>.

TOOL: {"name": "tb__deploy", "args": {"environment": "staging"}}

Plugin tools

Tools contributed by process plugins. Named plugin__<plugin>__<tool>.

TOOL: {"name": "plugin__security__scan", "args": {"path": "."}}

Skill-local MCP tools

Skills can activate bundled MCP servers when skills.allowMCP or the workspace MCP environment gate allows it. Skill-local tools use the same MCP naming and permission checks as other MCP tools.

Built-In Skills

Skills invoked via the skill tool. See AGENTS.md and Skills for full details.

Skill	Purpose
code_review	Review code changes for issues
bug_hunt	Systematically search for bugs
repo_summary	Generate repository overview
test_failure_analysis	Diagnose test failures
docs_writer	Draft manual-style documentation
commit_message	Write commit messages from diffs
ui_consistency_review	Check UI for consistency issues
security_review	STRIDE + OWASP Top 10 + LLM Top 10 + supply-chain review
simplify	Reuse, quality, and efficiency cleanup pass
browse_wiki	Search and read project docs/manual locally
full_output_enforcement	Generate complete code without placeholders/truncation
incident	Alert-link RCA runbook

Disable Tools

Disable specific tools globally via config.json:

{
  "tools": {
    "disabled": ["web_search", "mcp__*"]
  }
}

Values support glob patterns. Disabled tools are removed from the agent’s available tool set.

Keep Going

• Permissions - control what tools can do
• AGENTS.md and Skills - extend Anode with custom skills
• MCP - connect external tool servers
• Toolbox - add tools via shell scripts