The 14-tool catalog of a shipped open source AI agent

The top results for this keyword are framework roundups: LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Google ADK, Mastra, Dify, plus coding agents like OpenHands, Aider, and Cline. All of them are Python or TypeScript SDKs that a developer assembles into a product. None of them let a non-engineer install a working agent with a tool surface already wired in. Fazm is the other shape: a signed Mac app with 14 tools already declared in one file (acp-bridge/src/fazm-tools-stdio.ts, ALL_TOOLS at lines 279-503) and a Swift dispatcher that executes them (ChatToolExecutor.swift, lines 60-134). This page walks that catalog tool by tool so you can see what a shipped consumer agent's tool surface actually looks like, not another framework tutorial.

In source order: execute_sql, capture_screenshot, check_permission_status, request_permission, extract_browser_profile, edit_browser_profile, query_browser_profile, scan_files, set_user_preferences, ask_followup, complete_onboarding, save_knowledge_graph, save_observer_card, speak_response. They are declared in acp-bridge/src/fazm-tools-stdio.ts in the ALL_TOOLS array (lines 279-503) and dispatched in Desktop/Sources/Providers/ChatToolExecutor.swift (switch at lines 60-134). On top of that, the bundled MCP binary at /Contents/MacOS/mcp-server-macos-use exposes Accessibility-tree click and type tools (mcp__macos-use__*) that the agent can call to drive any other Mac app.

Screenshots force the LLM to solve a vision problem every turn (find the pixel, click the pixel) and they break when the target is occluded, offscreen, or rendered at a scaling factor the model was not trained on. Accessibility (AX) actuation treats the app as a tree of typed widgets and asks, 'find the AXButton named Send'. The reference returned is opaque and stable. You do not need coordinates. You do not need the window to be frontmost. You get AXEnabled, AXValue, AXSelectedText for free. Fazm's default path through the macos-use MCP is AX-first; capture_screenshot is explicitly documented as a fallback, not the default pipeline. That is why voice commands like 'mark the first reminder as done' land on the actual AXCheckBox, not a pixel guess.

Yes. fazm-tools-stdio.ts builds the TOOLS constant by filtering ALL_TOOLS (line 513-521). Chat observer sessions get only execute_sql, capture_screenshot, query_browser_profile, edit_browser_profile, and save_observer_card (CHAT_OBSERVER_TOOL_NAMES, line 271-277). Regular floating-bar sessions get the full set minus onboarding-only tools. Onboarding sessions get the full set. speak_response is filtered out entirely unless the voice toggle is on (VOICE_RESPONSE_TOOL_NAMES, line 506). So the tool list the LLM sees is always the intersection of 'what this session needs' and 'what the user enabled'.

The Claude agent speaks ACP (agent client protocol) to acp-bridge. acp-bridge spawns fazm-tools-stdio.ts as a subprocess (MCP server over stdio). When Claude emits tools/call, the stdio MCP forwards the call through a named pipe (FAZM_BRIDGE_PIPE env var) back to the parent acp-bridge process, which hands it to the Swift side of the app through WebRelay / ACPBridge. Swift dispatches in ChatToolExecutor.execute() and the return value walks back up the same pipe. For macos-use tools, acp-bridge spawns a separate MCP server, mcp-server-macos-use, that speaks directly to the Accessibility API. None of this is a cloud round trip; only the LLM call itself leaves the machine.

The guard is at two layers. First, ChatToolExecutor.swift blocks a keyword list of DROP, ALTER, CREATE, PRAGMA, ATTACH, DETACH, VACUUM outright (lines 152-154). Second, UPDATE and DELETE without a WHERE clause are rejected (lines 241-243). Multi-statement queries (semicolon-joined) are rejected (lines 186-191). The agent cannot issue a DROP TABLE even with prompt injection. SELECT is auto-limited to 200 rows. It is local-only, too; the database is fazm.db in Application Support, never synced to a server.

Yes, in two places. Add a block to ALL_TOOLS in acp-bridge/src/fazm-tools-stdio.ts with your tool's name, description, and JSON Schema. Then add a case in the switch in ChatToolExecutor.swift that implements the handler and returns a String. Rebuild and the LLM gets it on the next session. Because ALL_TOOLS is regular TypeScript, not a framework DSL, you can also add input validation, logging, or analytics wrappers in the dispatcher without fighting an orchestration layer.

Frameworks are libraries you use to build an agent; Fazm is the agent. LangGraph gives you stateful graph orchestration. CrewAI gives you multi-role agents with shared goals. AutoGen gives you a conversation-based agent loop. All three are correct choices if you are shipping your own voice assistant, your own coding agent, or your own research tool. Fazm is open-source source code for a finished consumer product. You can fork it, swap execute_sql for your own SQL engine, or replace capture_screenshot with whisper + diffusion. You cannot fork LangGraph and hand the result to your mom. Both shapes are valid. They are not the same artifact.

github.com/mediar-ai/fazm. The TypeScript declaration lives in acp-bridge/src/fazm-tools-stdio.ts; ALL_TOOLS starts at line 279 and ends at line 503. The Swift executor lives in Desktop/Sources/Providers/ChatToolExecutor.swift; execute() starts at line 57 and the switch runs lines 60-134. Every file path and line number on this page points to the public repo as of 2026-04-19.

Only for two things: the Claude LLM call itself, which is a commercial API round-trip, and Deepgram for speech-to-text when you use voice. All 14 tools run on-device. Browser profile extraction reads local browser SQLite files. execute_sql queries a local fazm.db. capture_screenshot takes a local screen capture. The macos-use MCP is a local binary. Fork the repo, swap the Claude call for Ollama or llama.cpp, and the whole stack is offline.