Tools

Overview

Tools are capabilities the agent can invoke during a conversation. Some tools require your approval before executing (marked with 🔒). Some tool schemas are deferred. The active harness exposes either load_tools or native tool_search as the loader and names the required call in the system prompt. Loading only makes a schema callable on the next model step; it does not execute the tool or bypass approvals.

System & files

Tool	Description	Approval
`bash`	Execute shell commands	🔒 Dangerous commands
`read_file` / `list_files` / `find_files` / `search_text`	Inspect local files without editing them	–
`write_file` / `edit_file`	Create or modify local files	🔒
`current_time`	Get current date and time	–
`research`	Spawn a research agent with all read-only tools (`depth`: quick/normal/deep)	–
`workflow`	Run a curated built-in multi-agent workflow	–
`background`	Spawn an autonomous background agent for long-running tasks	–
`render_html`	Render a sandboxed rich card in desktop chat (`display` or `input` mode)	–
`load_tools`	Load deferred tool schemas by group or exact name	–
`tool_search`	Find and load exact deferred tools in native-search runs	–
`request_connection`	Ask you to connect one required registered integration	–
`get_directives`	Read persistent behavioral directives	–
`set_directives`	Update persistent behavioral directives	🔒

File write tools and set_directives are deferred. The background tool can start new background work immediately; the background deferred group is only for listing, reading, or canceling existing background tasks.

Email

Requires Gmail to be connected. Deferred group: email.

Tool	Description	Approval
`emails`	Search and list emails	–
`read_email`	Read full email content	–
`send_email`	Send an email	🔒
`reply_email`	Reply in an existing thread by qualified `account:message_id`	🔒

Calendar

Requires Google Calendar to be connected. Deferred group: calendar.

Tool	Description	Approval
`calendar`	Search and list events	–
`create_calendar_event`	Create an event	🔒
`edit_calendar_event`	Edit an event	🔒
`delete_calendar_event`	Delete an event	🔒

Google Drive

Requires Google Drive to be connected. Deferred group: google_drive.

Tool	Description	Approval
`search_google_drive`	Search Docs and Sheets	–
`read_google_doc`	Read a Google Doc	–
`read_google_sheet`	Read a bounded A1 range	–
`create_google_doc` / `create_google_sheet`	Create an empty document or sheet	🔒
`edit_google_doc`	Append to or replace exact text	🔒
`update_google_sheet` / `append_google_sheet_rows`	Write cells or append rows	🔒

Facts

Facts are the canonical, source-backed correctness layer. Change them through an exact plan and commit; do not use a generic memory-record mutation tool.

Tool	Description	Approval
`search_facts`	Search visible canonical facts; returns stable `fact_id` values.	–
`get_fact`	Read one fact, including lifecycle and saved provenance.	–
`get_fact_history`	Read a fact’s immutable event history and provenance.	–
`get_due_fact_reviews`	List facts due for evidence-based review.	–
`plan_fact_changes`	Validate creates, reviews, amendments, supersessions, expiries, or retractions without publishing them.	–
`commit_fact_changes`	Commit one exact returned plan if its facts remain unchanged.	–

plan_fact_changes returns the preview and plan_id; commit_fact_changes accepts only that plan_id. The server derives authority and provenance from the active session and Area.

Wiki

Wiki pages are the readable, managed projection. Read before writing, then use semantic page arguments; the backend enforces concurrency. On a conflict, reread and retry. Scheduled automations can use generic writes only below automations/; generated publishing remains reserved for an explicit owner. This group is deferred: load it with load_tools(group="wiki").

Tool	Description	Approval
`list_wiki_pages`	List active pages and direct child directories in one managed directory.	–
`read_wiki_page`	Read one active page by exact managed path.	–
`wiki_links`	Read a page’s outgoing links and backlinks.	–
`create_wiki_page`	Create one common page from its path, title, and body.	🔒
`edit_wiki_page`	Replace one page body by its exact managed path.	🔒
`archive_wiki_page`	Recoverably archive one page by its exact managed path.	🔒
`move_wiki_page`	Move one page by its exact managed path and rewrite its wiki links atomically.	🔒
`publish_wiki_generated`	Update one automation-owned section without touching the rest of its page.	🔒

Web

Web search uses WEB_SEARCH mode (auto, exa, ddgs, none). EXA_API_KEY is only required when mode resolves to Exa.

Tool	Description	Approval
`web_search`	Search the web	–
`web_fetch`	Fetch content from a URL	–

Slack

Requires Slack to be connected. Deferred group: slack.

Tool	Description	Approval
`slack_search`	Search Slack messages	–
`slack_channel`	Read recent channel history	–
`slack_thread`	Read a message thread	–
`slack_channels`	List accessible channels	–
`slack_dms`	List open direct messages	–
`slack_dm`	Read a direct message conversation	–
`slack_users`	Search workspace members	–
`slack_user`	Read a user profile	–
`slack_file`	Read a Slack file or image by its returned reference	–
`slack_post_message`	Post a message	🔒
`slack_post_blocks`	Post Block Kit content	🔒

Automations

Deferred group: automations.

Tool	Description	Approval
`create_automation`	Schedule a task	🔒
`create_loop`	Start an in-chat recurring loop	🔒
`schedule_wakeup`	Schedule a one-off wakeup into the current session	🔒
`loop_done`	Stop the current loop after its completion condition is met	🔒
`list_automations`	List all automations	–
`update_automation`	Modify an automation	🔒
`delete_automation`	Remove an automation	🔒
`run_automation`	Trigger immediate run	🔒
`get_automation_result`	Get last run result	–

Runtime state

Tool	Description	Approval
`update_todos`	Update the visible task list for complex work	–
`get_goal`	Read the active durable session goal	–
`complete_goal`	Mark the active goal complete after verification	–
`block_goal`	Mark the active goal blocked on missing input	–
`create_session`	Create a new chat/channel session	–
`list_recent_sessions` / `read_session` / `search_transcripts`	Inspect readable chat history	–
`send_message`	Message a chat, or steer an agent you spawned	🔒 (chats only)

App control

Available only when the active session has app-control capability.

Tool	Description	Approval
`rename_session`	Rename a chat or agent session	🔒
`archive_session`	Archive finished sessions	🔒
`request_attention`	Add a needs-you item to Home	🔒
`open_in_app`	Open a supported resource in the desktop app	🔒

These are mostly internal coordination tools. The desktop app exposes the visible results as todos, channels, session history, and status updates.

Background tasks

Deferred group: background.

Tool	Description	Approval
`cancel_agent`	Stop an agent you spawned, by its session id	🔒

Agents are spawned with the always-available background tool, which returns the agent’s own session id. Results are delivered back into the parent chat automatically; use read_session to inspect an agent’s work and send_message to steer it.

Notifications

Deferred group: notifications.

Tool	Description	Approval
`notify`	Send a notification	🔒

MCP tools

Connected MCP server tools are deferred by server. Use load_tools(group="mcp:<server>"), for example load_tools(group="mcp:obsidian"). MCP tools still keep their own approval/mutation behavior after loading. MCP tools default to conservative external execution: approval required. Per-tool config can override that with tool_policies.<tool_name>. If a server is trusted, trust_tool_annotations=true lets Arden use MCP Tool.annotations hints such as readOnlyHint and destructiveHint; explicit tool_policies still win.

Canonical workflows

Follow the same sequence for source-of-truth work: search → inspect → preview → mutate → verify. Never invent IDs; pass stable refs returned by the previous tool.

Bounded file edit

Find candidates with find_files or search_text; if has_more is true, narrow the query or continue with its cursor.
read_file(path, offset, limit) the exact region and retain its revision.
Call edit_file or write_file with that revision. The approval card shows the diff before the write.
Treat write_conflict, not_found, and permission_denied as terminal for that attempt; follow recovery_action, re-read, and create a new preview.
Verify the returned after_ref by reading the file again. Large results provide a durable raw_ref rather than disappearing at truncation.

Slack and Gmail mutation

Load slack or email, search with a bounded limit, and continue only when the result says more may exist.
Inspect the selected message/thread. Use its exact source_refs[].ref (channel:timestamp for Slack, account:message_id for Gmail).
Draft the post, send, or reply; provide a stable idempotency_key. Approval shows the destination and bounded body preview.
A successful mutation returns an effect, provider receipt, and after_ref; an uncertain result says to verify provider state before retrying.
Read the returned ref to verify. For Gmail replies, reply_email preserves In-Reply-To, References, and Gmail threadId.

MCP recovery

Load the exact server group with load_tools(group="mcp:<server>"); loading never executes a tool.
Inspect the lossless schema and annotations. Bound list arguments and use refs from earlier results.
Preview and approve external mutations unless an explicit policy says otherwise.
On tool_not_loaded, load the named group; on mcp_error, follow its recovery action. Do not retry uncertain mutations without checking state.
Preserve returned source_refs, raw_ref, receipts, and verification data for downstream calls.

Deferred loading examples

Need	Load call
Search Gmail	`load_tools(group="email")`
Read or edit calendar	`load_tools(group="calendar")`
Search Slack	`load_tools(group="slack")`
Browse or publish managed wiki pages	`load_tools(group="wiki")`
Manage automations	`load_tools(group="automations")`
Spawn or inspect background work	`load_tools(group="background")`
Send a notification	`load_tools(group="notifications")`
Update standing directives	`load_tools(group="directives")`
Use an MCP server	`load_tools(group="mcp:<server>")`

Skills

Tool	Description	Approval
`use_skill`	Activate a skill for specialized instructions	–
`create_skill`	Save an approved reusable global skill	🔒

Obsidian should be connected as an MCP server and loaded with load_tools(group="mcp:obsidian") when needed.

Approval flow

When a tool requires approval, the desktop app shows an approval card with the tool name, arguments, preview, and options to approve or reject the call. Session/tool-specific policy overrides can skip repeated approvals when you explicitly configure them. Settings → Tools can override individual tools:

Override	Behavior
Default	Use the tool’s policy
Approve	Expose the tool and skip approval
Ask	Expose the tool and require approval
Deny	Hide the tool from the agent and block execution

Per-automation tool scoping

An automation can carry a tool scope: an allowlist of tool-name patterns that bounds what its runs may touch, applied as a hard outer gate before any other selection. Patterns are * (everything), an exact name, or a prefix* wildcard.

"tool_scope": ["search_facts", "get_fact", "get_fact_history", "list_wiki_pages", "read_wiki_page", "wiki_links", "web_search"]

This is how area agents stay observe-only: they get bounded fact and wiki reads, but no fact commits, generated-page publishing, or external actions. A missing scope gives an automation only the read-only floor. Use "*" only when unrestricted tool access is intentional. Area custodians finish by calling submit_area_report. Invalid or stale reports return a tool error inside the same agent loop, so the custodian can correct the report before the automation is allowed to settle.

User-defined tools

Create custom tools in ~/.arden/tools/. Each file exports a tools mapping. Inputs must be bounded, failures typed, and stable results sourced:

from pydantic import BaseModel, Field

from arden.agent.types.tools import ToolSourceRef
from arden.tools.core import ToolAction, ToolPolicy, ToolResult, ToolScope, tool
from arden.tools.core.context import ToolExecution


class LookupInput(BaseModel):
    query: str = Field(min_length=1, max_length=500)
    limit: int = Field(default=20, ge=1, le=100)


async def lookup(execution: ToolExecution, args: LookupInput) -> ToolResult:
    service = execution.ctx.services.get("catalog")
    if service is None:
        return ToolResult.failure(
            code="not_configured",
            message="Catalog is unavailable.",
            preview="Unavailable",
            recovery_action="Configure the catalog service and retry.",
        )
    rows = await service.search(args.query, limit=args.limit)
    return ToolResult(
        content="\n".join(row.title for row in rows),
        preview=f"{len(rows)} results",
        source_refs=tuple(
            ToolSourceRef(provider="catalog", kind="item", ref=row.id, title=row.title)
            for row in rows
        ),
    )


tools = {
    "lookup": tool(
        display_name="Lookup",
        description="Search the configured catalog and return stable item refs.",
        input_model=LookupInput,
        policy=ToolPolicy(
            action=ToolAction.READ,
            scope=ToolScope.INTERNAL,
            permissions=frozenset({"catalog"}),
        ),
        execute=lookup,
    )
}

Use policy=ToolPolicy(..., requires_approval=True) plus an approval= function for tools that change source-of-truth state. Use permissions=frozenset({...}) to hide a tool unless a service is configured. Restart the server after adding or changing user tools.

Getting Started

Features

Integrations

Reference

Overview

System & files

Email

Calendar

Google Drive

Facts

Wiki

Web

Slack

Automations

Runtime state

App control

Background tasks

Notifications

MCP tools

Canonical workflows

Bounded file edit

Slack and Gmail mutation

MCP recovery

Deferred loading examples

Skills

Approval flow

Per-automation tool scoping

User-defined tools

​Overview

​System & files

​Email

​Calendar

​Google Drive

​Facts

​Wiki

​Web

​Slack

​Automations

​Runtime state

​App control

​Background tasks

​Notifications

​MCP tools

​Canonical workflows

​Bounded file edit

​Slack and Gmail mutation

​MCP recovery

​Deferred loading examples

​Skills

​Approval flow

​Per-automation tool scoping

​User-defined tools

Overview

System & files

Email

Calendar

Google Drive

Facts

Wiki

Web

Slack

Automations

Runtime state

App control

Background tasks

Notifications

MCP tools

Canonical workflows

Bounded file edit

Slack and Gmail mutation

MCP recovery

Deferred loading examples

Skills

Approval flow

Per-automation tool scoping

User-defined tools