AgentWorks: The AI Agent Platform That Chases Goals

Who It's For

AgentWorks is officially positioned as a source-available platform for goal-driven AI agents: you set an agent a target metric, and it keeps working until that metric is reached. If you already use coding-agent CLIs such as Claude Code, Codex, Cursor, Pi, or Muse, it runs directly on top of those existing CLIs.

It is also an MCP server: if you want to check goals, read reports, and start runs directly from chats in ChatGPT, Claude (Desktop, Code, Cowork), or any MCP client, it is a strong fit.

Prerequisites worth noting: the officially documented install path covers macOS only (installs into /Applications, clears the macOS quarantine flag, and installs Go through Homebrew if needed); other operating systems are not mentioned. It is not an out-of-the-box hosted service either — on first launch you must pick your own local workspace folder and set an AUTH_SECRET.

Your First Try in 3 Minutes

Prerequisites: a macOS machine and a terminal; the script installs Go through Homebrew when needed.

Run the official one-line install command in a terminal:

curl -fsSL https://raw.githubusercontent.com/manishiitg/coding-agent-loop/main/install.sh | bash

Before pressing Enter, understand the side effects: the script downloads the latest release, installs AgentWorks.app into /Applications, installs the MCP bridge that Claude Code and Codex use for tool access (placed in ~/go/bin), clears the macOS quarantine flag, and then launches the app. This is an install action with local writes, not a read-only operation; if you have concerns about remote scripts, view the URL's content in a browser first.

First-launch configuration: follow the app's prompts to choose a workspace folder and set an AUTH_SECRET. It is used to encrypt provider keys; if the same workspace is opened on multiple machines, every machine must reuse the same value.

A low-risk first call: the install script has already set up the MCP bridge for Claude Code and Codex, so from Claude Code ask it to check goals or read reports — these are the read-type capabilities described officially and do not trigger runs. Success signal: Claude Code returns goal status or report content. Integration steps for other MCP clients (such as ChatGPT or Claude Desktop) are not shown on the official page and are not invented here.

Use Cases

Case 1: Check progress and read reports in chat. Scenario: you have already handed a metric-chasing goal to AgentWorks and want to know how it is going from your everyday conversation. How: in a connected client, ask in natural language to check goals and read reports (check goals, read reports). Expected result: the official documentation explicitly supports viewing goals and reading reports from that chat; specific return fields are not listed, so go by the actual output.

Case 2: Start a run from chat. Scenario: you decide a goal needs to move forward right now and want to kick it off without leaving the current conversation. How: ask AgentWorks to start runs. Expected result: the run starts and is carried by the coding-agent CLIs you already have installed (Claude Code, Codex, Cursor, Pi, Muse). Note that this is not a query but an action that does real work — the run keeps working until the metric is reached. Confirm the goal and scope configuration before starting; run-level human confirmation checkpoints are not documented officially.

Usage Boundaries

Read/write nature: the capabilities exposed over MCP include both reads (check goals, read reports) and execution (start runs). Starting a run makes the agent keep working until the target metric is reached, with real side effects; whether individual tools carry a read-only hint is not labeled officially, so do not assume everything is read-only.

Credentials and local data: first launch requires choosing a local workspace folder; AUTH_SECRET is used to encrypt provider keys, and the same workspace must reuse the same value on every machine. How provider keys are configured is not elaborated on the official page.

Side effects and confirmation points: the install script writes to /Applications and ~/go/bin, clears the macOS quarantine flag, and launches the app; once a run starts, the agent keeps executing, and no step-by-step confirmation mechanism was found in the official documentation.

Explicit limits and unknowns: the install path is described for macOS only; the platform is source-available (the repository title also calls itself Open source), and license terms are governed by the repository; non-macOS platform support, the complete MCP tool list and parameters, remote hosting options, and authentication methods other than AUTH_SECRET are all undocumented — verify against the official repository before use.