Features Overview

EdgeCrab ships as a single static binary with enterprise-grade features. No Python venv, no Node.js — just one executable.

Input --> ContextBuilder --> AgentLoop --> ToolRegistry --> Security checks
                                 ^                              |
                                 +---- ToolResult <------------+

---

Core Features

Autonomous ReAct Loop

EdgeCrab runs a Reason-Act-Observe loop — it reasons about a task, calls a tool, observes the result, then repeats. The loop runs up to model.max_iterations tool calls (default: 90) before stopping.

For weaker models or long tool-heavy runs, EdgeCrab can also enable an opt-in Shadow Judge completion oracle. It performs a cheap secondary LLM verdict before the loop accepts a final answer, helping catch premature “I’m done” stops that still have missing sub-steps.

Ratatui TUI

A full-featured terminal UI with:

• Streaming token display with cost tracking
• Tool execution feed with per-tool timing
• Slash command autocomplete (all installed skills appear as commands)
• Keyboard-driven interface
• Customizable skins, symbols, and personality presets
• Mission Steering (Ctrl+S) — inject hints, redirects, or stop signals into a running agent loop mid-turn
• Shadow Judge controls (/shadow-judge) — toggle the completion oracle per session and see intervention badges when it keeps the run alive

See TUI Interface and Mission Steering

15 LLM Providers

Switch provider and model without restarting:

/model openai/gpt-5
/model openai/gpt-4o
/model ollama/llama3.3
/model copilot/gpt-4.1-mini

See LLM Providers

Skills System

Reusable Markdown workflows that teach EdgeCrab domain-specific tasks:

• Skills are directories containing a SKILL.md file
• EdgeCrab can create and improve skills during sessions
• Compatible with agentskills.io

See Skills System

Persistent Memory

Agent memory stored in ~/.edgecrab/memories/:

• Auto-written after each session when memory.auto_flush: true
• Injected into the system prompt at session start
• Honcho integration for cross-session user modeling

See Memory

Browser Automation

Built-in browser control via Chrome DevTools Protocol (CDP):

• Navigate, click, type, scroll, take screenshots
• Console log capture
• Vision analysis of screenshots
• Session recording as WebM

See Browser Automation

Semantic Coding with LSP

Built-in language-server support gives EdgeCrab semantic code intelligence instead of text-only heuristics:

• Go to definition, references, implementation, document symbols, and workspace symbols
• Code actions, semantic rename, whole-document and range formatting
• Inlay hints, semantic tokens, signature help, call hierarchy, and type hierarchy
• Pull diagnostics, workspace type-error scans, linked editing, and LLM-enriched diagnostic explanations

See Language Server Protocol

15 Messaging Gateways

Run EdgeCrab as a persistent bot on 15 platforms:

• Telegram, Discord, Slack, Signal, WhatsApp
• Matrix, Mattermost, DingTalk, SMS, Email
• Home Assistant, Webhook, API Server, Feishu/Lark, WeCom

See Messaging Gateway

Security

Built-in defense in depth (7 layers):

• Path traversal prevention (SanitizedPath compile-time type)
• SSRF guard with private-IP blocklist (SafeUrl compile-time type)
• Aho-Corasick command scanner (8 danger categories)
• Prompt injection detection in context files
• Code execution sandbox (API keys stripped from child env)
• Skills threat scanner (23 patterns — exfiltration, injection, persistence)
• Output redaction (API keys, tokens)

See Security Model

Checkpoints and Rollback

Shadow git commits before every destructive file operation. Roll back at any time:

/rollback       # interactive checkpoint browser

See Checkpoints

Cron / Scheduled Tasks

Built-in cron scheduler with agent-managed jobs:

edgecrab cron add "0 9 * * 1-5" "morning standup summary"

See Cron Jobs

---

Tool Inventory

All tools sourced from CORE_TOOLS constant in crates/edgecrab-tools/src/toolsets.rs.

Toolset Aliases

Alias	Expands to
core	file + meta + scheduling + delegation + code_execution + lsp + session + mcp + messaging + media + browser (runtime-gated)
coding	file + terminal + search + code_execution + lsp
research	web + browser + vision
debugging	terminal + web + file
safe	web + vision + image_gen + moa
minimal	file + terminal
data_gen	file + terminal + web + code_execution
all	every registered tool (no filtering)

File Tools (file)

Tool	Description
read_file	Read file contents with optional line range
write_file	Write or overwrite a file (creates checkpoint)
patch	Apply a unified diff patch to a file (creates checkpoint)
search_files	Regex or glob search across file tree

Terminal Tools (terminal)

Tool	Description
terminal	Run a shell command and capture output
run_process	Start a long-running background process
list_processes	List running background processes
kill_process	Kill a process by ID
get_process_output	Get stdout/stderr from a background process
wait_for_process	Block until a process exits
write_stdin	Send input to a process’s stdin

Web Tools (web)

Tool	Description
web_search	DuckDuckGo search (SSRF-guarded)
web_extract	Extract text content from a URL
web_crawl	Recursive site crawl with optional depth limit

Browser Tools (browser)

Runtime-gated: requires Chrome or Chromium. If no browser is reachable, these tools are silently absent from the tool list.

Tool	Description
browser_navigate	Navigate to a URL
browser_snapshot	Get page accessibility tree as text
browser_click	Click an element
browser_type	Type text into an input
browser_scroll	Scroll the page
browser_press	Press a keyboard key
browser_back	Navigate back
browser_close	Close the browser
browser_console	Capture browser console logs
browser_get_images	Get images from the page
browser_vision	Take screenshot and analyze with vision model
browser_wait_for	Wait for element/text to appear
browser_select	Select a dropdown option
browser_hover	Hover to trigger tooltips/states

Memory Tools (memory)

Tool	Description
memory_read	Read a memory file
memory_write	Write or update a memory file
honcho_conclude	Commit a Honcho cross-session memory entry
honcho_search	Search the Honcho user model
honcho_list	List Honcho memory entries
honcho_remove	Remove a Honcho entry
honcho_profile	Update the Honcho user profile
honcho_context	Get relevant Honcho context for current task

Skills Tools (skills)

Tool	Description
skills_list	List available skills
skills_categories	List skill categories
skill_view	View a skill’s content
skill_manage	Install / uninstall / update a skill
skills_hub	Browse the agentskills.io hub

Scheduling Tools (scheduling)

Tool	Description
manage_cron_jobs	Create / list / delete / enable / disable cron jobs

Meta Tools (meta)

Tool	Description
manage_todo_list	Create and track a session todo list
clarify	Ask the user a clarifying question (TUI and gateway only)

Delegation Tools (delegation)

Tool	Description
delegate_task	Spawn a subagent for a parallel subtask
moa	Multi-model consensus reasoning

Code Execution (code_execution)

Tool	Description
execute_code	Execute Python, Node.js, or Bash code in isolation

LSP Tools (lsp)

Tool	Description
lsp_goto_definition	Jump to a symbol definition
lsp_find_references	Find symbol references
lsp_hover	Get hover docs and type information
lsp_document_symbols	Enumerate symbols in a file
lsp_workspace_symbols	Search symbols across the workspace
lsp_goto_implementation	Jump to concrete implementations
lsp_call_hierarchy_prepare	Prepare call hierarchy items
lsp_incoming_calls	Show incoming calls
lsp_outgoing_calls	Show outgoing calls
lsp_code_actions	List server-suggested fixes and refactors
lsp_apply_code_action	Resolve and apply a code action
lsp_rename	Rename a symbol across files
lsp_format_document	Format a whole file
lsp_format_range	Format a selected range
lsp_inlay_hints	Return inlay hints
lsp_semantic_tokens	Return semantic token classes
lsp_signature_help	Show function signature help
lsp_type_hierarchy_prepare	Prepare a type hierarchy item
lsp_supertypes	List supertypes
lsp_subtypes	List subtypes
lsp_diagnostics_pull	Pull document or workspace diagnostics
lsp_linked_editing_range	Return linked editing regions
lsp_enrich_diagnostics	Explain diagnostics with the LLM
lsp_select_and_apply_action	Pick and apply the best action
lsp_workspace_type_errors	Summarize workspace-wide type errors

Session Tools (session)

Tool	Description
session_search	Full-text search (FTS5) across all session history

MCP Tools (mcp)

Tool	Description
mcp_list_tools	List tools from connected MCP servers
mcp_call_tool	Call a tool on an MCP server
mcp_list_resources	List resources from MCP servers
mcp_read_resource	Read a resource from an MCP server
mcp_list_prompts	List prompts from MCP servers
mcp_get_prompt	Get a prompt from an MCP server

Media Tools (media)

Tool	Description
text_to_speech	Convert text to speech (edge-tts, OpenAI, ElevenLabs)
vision_analyze	Analyze an image file with a vision model
transcribe_audio	Transcribe an audio file with Whisper
generate_image	Generate an image (runtime-gated: requires provider support)

Home Assistant Tools (homeassistant)

Runtime-gated: requires HA_URL and HA_TOKEN environment variables.

Tool	Description
ha_list_entities	List all Home Assistant entities
ha_get_state	Get current state of an entity
ha_list_services	List available Home Assistant services
ha_call_service	Call a Home Assistant service

Core Tools (core)

Tool	Description
checkpoint	Create / list / restore / diff filesystem checkpoints

Messaging Tools (runtime-gated)

Only available in gateway sessions (Telegram, Discord, Slack, etc.):

Tool	Description
send_message	Send a message to the current chat (gateway only)

---

What’s Next?

• ReAct Tool Loop — How the autonomous reasoning engine works
• TUI Interface — Full keyboard shortcuts and slash commands
• Skills System — Creating and using reusable skills
• Memory — Persistent memory and Honcho user modeling
• Browser Automation — Browser automation with CDP
• Language Server Protocol — Semantic code navigation, edits, and diagnostics
• SQLite State and Search — Session persistence and FTS5 search
• Security Model — 7-layer defense stack

---

Feature Coverage at a Glance

User Input
    |
    v
[Context Builder]  <-- SOUL.md + AGENTS.md + memories + skills
    |
    v
[Agent Loop]       -- LLM reasons, calls tools, observes results
    |
    +--[Security checks]  <-- path jail + SSRF + command scan + approval
    |
    +--[Tool Dispatch]    <-- file / terminal / web / browser / lsp / memory / mcp
    |
    v
[State DB (SQLite)]  -- every message stored, FTS5 indexed
    |
    v
[TUI / Gateway]     -- rendered in ratatui or sent to Telegram/Discord/...

---

Frequently Asked Questions

Q: What’s the difference between “tools” and “skills”?

Tools are atomic Rust functions (read a file, run a command, search the web). Skills are Markdown documents that guide how the agent uses tools. A skill says “when doing a security audit, first do X, then Y, then Z.” Tools are the verbs; skills are the recipe.

Q: How does EdgeCrab compare to agentic frameworks like LangChain?

EdgeCrab is a self-contained binary, not a framework. It’s designed to be run, not extended programmatically (though SDKs exist). You configure it via YAML and Markdown, not code. For programmatic use, the Python/Node.js SDKs provide an API surface.

Q: Can EdgeCrab run without internet access?

Yes. Use --model ollama/llama3.3 (local Ollama) and --toolset file,terminal,memory. All tools used by those toolsets work offline.

Q: How many tools can be active at once?

All registered tools can be active simultaneously, including the full LSP surface when semantic coding is enabled. Limiting toolsets to what’s needed keeps the system prompt shorter and the LLM more focused.

Q: Edge case: Can the agent call the same tool infinitely?

No. model.max_iterations (default: 90) limits total tool calls per session. The LLM also receives tool results that typically converge toward an answer. If the budget is exhausted, EdgeCrab reports to the user.