Skip to content

Latest commit

 

History

History
198 lines (128 loc) · 5.97 KB

File metadata and controls

198 lines (128 loc) · 5.97 KB

Tools Guide

Built-in tools that the agent can call (via function/tool calling).

Tool Safety Model

  • Core tool toggles live in preferences/runtime-config.json (dashboard: Settings).
  • Official integrations such as browser, Brave Search, Tavily Search, Firecrawl, Perplexity Sonar, weather, and mail are loaded through the plugin runtime and keep their own configuration under preferences/plugins/<owner>/<plugin>.json.
  • File and shell access are sandboxed to the tool workspace (bot.tools.*.workspace).
  • Destructive actions may require user confirmation when tool confirmations are enabled.

Runtime Configuration

Core tools are controlled via runtime-config.json under tools and security:

{
  "tools": {
    "filesystemEnabled": true,
    "shellEnabled": true,
    "skillManagementEnabled": true,
    "skillTransitionEnabled": true,
    "tierEnabled": true,
    "goalManagementEnabled": true,
    "shellEnvironmentVariables": []
  },
  "security": {
    "toolConfirmationEnabled": false,
    "toolConfirmationTimeoutSeconds": 60
  }
}

Plugin-owned tool settings are stored separately, for example:

  • preferences/plugins/golemcore/browser.json
  • preferences/plugins/golemcore/brave-search.json
  • preferences/plugins/golemcore/tavily-search.json
  • preferences/plugins/golemcore/firecrawl.json
  • preferences/plugins/golemcore/perplexity-sonar.json
  • preferences/plugins/golemcore/weather.json
  • preferences/plugins/golemcore/mail.json

Workspace Sandboxing (FileSystem/Shell)

File and shell tools operate only inside the configured workspace:

  • bot.tools.filesystem.workspace
  • bot.tools.shell.workspace

In Docker, set these via env and mount volumes:

docker run -d \
  -e STORAGE_PATH=/app/workspace \
  -e TOOLS_WORKSPACE=/app/sandbox \
  -v golemcore-bot-data:/app/workspace \
  -v golemcore-bot-sandbox:/app/sandbox \
  -p 8080:8080 \
  golemcore-bot:latest

Built-in Tools

Dashboard IDE API vs filesystem Tool

The dashboard embedded IDE uses direct authenticated HTTP endpoints under /api/files. This API is separate from the filesystem tool used by the agent tool loop.

  • Dashboard IDE API (/api/files) is deterministic UI/backend integration.
  • filesystem tool is an LLM-callable tool with tool-loop semantics.

Both are sandboxed to the configured filesystem workspace root.

filesystem

Sandboxed file operations (read/write/list/create/delete/send) under the tool workspace.

Operations:

  • read_file, write_file, list_directory, create_directory, delete, file_info, send_file

shell

Sandboxed command execution under the tool workspace.

Input fields:

  • command (required)
  • timeout (seconds; default/max enforced)
  • workdir (relative to tool workspace)

browse

Headless browsing (Playwright) to extract page text, html, or capture screenshot. Provided by the official golemcore/browser plugin.

Input fields:

  • url (required)
  • mode: text | html | screenshot

brave_search

Web search via Brave Search API.

Provided by the official golemcore/brave-search plugin. Configure it in Settings or preferences/plugins/golemcore/brave-search.json.

tavily_search

Web search via Tavily Search API with optional answer synthesis and advanced retrieval depth.

Provided by the official golemcore/tavily-search plugin. Configure it in Settings or preferences/plugins/golemcore/tavily-search.json.

firecrawl_scrape

Remote webpage scraping via Firecrawl /v2/scrape.

Provided by the official golemcore/firecrawl plugin. Configure it in Settings or preferences/plugins/golemcore/firecrawl.json.

perplexity_ask

Grounded synchronous completions via Perplexity Sonar with selectable search model.

Provided by the official golemcore/perplexity-sonar plugin. Configure it in Settings or preferences/plugins/golemcore/perplexity-sonar.json.

skill_management

Create/list/get/delete skills at runtime (writes under workspace/skills/ in the storage workspace).

skill_transition

Switch active skill/pipeline step.

set_tier

Switch model tier mid-conversation (balanced, smart, coding, deep).

goal_management

Auto Mode goal/task/diary management.

plan_get / plan_set_content

Plan-work tools for canonical plan drafting (SSOT).

  • Advertised only when plan mode is active for the current session.
  • plan_get returns the current canonical Markdown plan draft.
  • plan_set_content accepts full Markdown in plan_markdown (optional title) and requests finalization.
  • Persistence/finalization is handled by PlanFinalizationSystem; plan_set_content is a deterministic control signal.
  • Finalizing a draft makes the plan READY, while plan work remains active until /plan done or /plan off.

memory

Structured memory operations for autonomous workflows.

Operations:

  • memory_search: search memory by query
  • memory_read: read specific items by id, fingerprint, or query
  • memory_expand_section: ask for more detail from a compact prompt section such as semantic facts or recent episodes
  • memory_add: write a semantic/procedural memory item
  • memory_update: update an existing semantic/procedural memory item
  • memory_promote: promote matching memories into a longer-lived layer
  • memory_forget: forget matching memories by id, fingerprint, or query

This tool uses Memory V2 APIs (no direct filesystem assumptions for memory writes).

When progressive disclosure is enabled, memory_read and memory_expand_section are the main "show me more" operations.

imap / smtp

Email tools.

Provided by the official golemcore/mail plugin. Configured in Settings or preferences/plugins/golemcore/mail.json.

send_voice

Synthesize and send voice responses (when voice is enabled and configured).

datetime

Built-in utility tool.

weather

Weather lookup provided by the official golemcore/weather plugin.

See Also