AgentWorks

Give an AI agent a goal and a metric. It keeps working until it hits the target.

AgentWorks is an open-source platform for goal-driven AI agents. You describe an outcome, pick the number that proves it, and set a target. Agents plan the work, run it on a schedule, measure every run, and change their own plan until the metric moves. It runs on the coding-agent CLIs you already use: Claude Code, Codex, Cursor, Pi, and Muse.

Website · Docs · First workflow · Releases · Book a call

_{Illustrative example.}

How it works

Goals	Every workflow has a goal: the outcome in plain words, a primary metric with a target, supporting metrics, and rules that must stay true. Progress is shown as dated measurements. Missing or stale numbers are flagged, never guessed. Goal measurement
Auto-improve	After runs, AgentWorks reviews the evidence against the goal. It repairs broken steps, drops approaches that didn't move the metric, does the work nobody was doing, and checks later whether it helped. Improvement system · Auto-improvement framework
Autonomy	You choose how far it goes, per workflow: Ask first, Run steps, Edit workflow, or Full. Anything outward-facing can always require your approval.
Crew	Always-on teammates, each with its own files, tools, memory, browser, and schedule. Talk to them in the app, Slack, or WhatsApp. Bot connectors
Playbooks	Ready-made agent setups you install into a workflow: 23 today for browser QA, reliability, security, performance, FinOps, and growth analytics. Plain Markdown skill packages, so you can write your own. Playbooks

Install (macOS, Apple Silicon)

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

This downloads the latest release, installs AgentWorks.app to /Applications, installs the MCP bridge that Claude Code and Codex use for tool access (to ~/go/bin, installing Go through Homebrew if needed), clears the macOS quarantine flag, and launches the app. Pin a version with RUNLOOP_VERSION=v1.25.6 curl -fsSL … | bash (the variable keeps its legacy name).

On first launch, pick a workspace folder and set an AUTH_SECRET (used to encrypt provider keys; reuse the same value on every machine that opens this workspace). Then connect a coding-agent CLI in LLM Configuration, signing in with your subscription or that CLI's API key.

Manual install, and the "AgentWorks is damaged and can't be opened" message

Download AgentWorks-<version>-arm64.dmg from the latest release and drag the app to Applications.

The build is not yet signed or notarized, so macOS Gatekeeper flags it after download. The app is fine; clear the quarantine flag:

xattr -cr /Applications/AgentWorks.app

If macOS still complains, also run xattr -cr ~/Downloads/AgentWorks-*.dmg. No sudo is needed. Releases installed under the old name use /Applications/Runloop.app. Signing and notarization are on the roadmap.

Works with

• Coding-agent CLIs: Claude Code, OpenAI Codex CLI, Cursor CLI, Pi CLI (Gemini, OpenRouter, and other Pi providers), and Muse. Use the subscription you already pay for.
• Channels: Slack and WhatsApp for two-way conversations; Gmail for outbound updates.
• Tools: any MCP server, workspace files, and a persistent, isolated browser per workflow (browser docs).
• Your AI app: AgentWorks is also an MCP server. Connect ChatGPT, Claude (Desktop, Code, Cowork), or any MCP client to check goals, read reports, and start runs from that chat. It can read and run, but not edit workflows (CLI and MCP).

AgentWorks runs each vendor's own agent (model plus harness), not a bare model API, so you get everything the agent can do. Route each step to the agent and model that fit it: your strongest for judgment, a cheaper one for routine steps.

Control and security

• Approvals: agents prepare changes and wait for approve, reject, or defer (human feedback).
• Secrets vault: credentials are encrypted and injected only at run time, never shown in chat or logs (secrets).
• Sandboxing: OS-enforced per agent, Landlock on Linux and sandbox-exec on macOS, limited to the folders and tools each workflow is granted (FolderGuard).
• Run logs and cost: every step, tool call, and model cost is recorded per run and per workflow (cost and logs).
• Accounts: admin, member, contributor, and read-only roles, with owner or reader access per workflow (multi-user).

Run from source

Prerequisites: Go 1.26+, Node.js 20+, and whichever coding-agent CLIs you want to use.

The backend builds against two engine libraries, mcpagent and multi-llm-provider-go, which agent_go/go.mod expects to find next to this repo. Clone all three side by side:

mkdir agentworks && cd agentworks
git clone https://github.com/manishiitg/coding-agent-loop.git
git clone https://github.com/manishiitg/mcpagent.git
git clone https://github.com/manishiitg/llm-provider-mcp.git multi-llm-provider-go

cd coding-agent-loop
(cd frontend && npm ci)
(cd agent_go && go mod download)
./run_agentworks

./run_agentworks starts the agent API, workspace API, frontend, and Electron shell. No .env is needed for a local run; the launcher creates agent_go/.env with a persistent AUTH_SECRET. Don't copy agent_go/env.example for local use, since it is for managed deployments.

Service	Default URL
Agent API	http://localhost:18743
Workspace API	http://localhost:18744
Frontend	http://127.0.0.1:51733

If a port is busy, the runner picks the next free one and prints it. Logs go to agent_go/logs/. Run ./run_agentworks --help for options such as --only-frontend and --build.

Checks before a pull request:

(cd agent_go && go test ./cmd/server -run '^$')   # backend compiles
(cd frontend && ./node_modules/.bin/tsc -b)       # frontend type-checks
./scripts/scan-secrets.sh                         # no secrets committed

Install the git hooks once with ./scripts/install-git-hooks.sh.

Self-hosting

Run AgentWorks on your Mac, or on your own Linux server with the rootless deployer. See the deployment overview. For a managed or private-cloud deployment with SSO, audit logs, and support, talk to us.

Contributing

Issues and pull requests are welcome. Start with the docs index and the workflow system overview. Keep changes focused and run the checks above.

License

MIT.

Built on the Model Context Protocol, our open-source engine libraries mcpagent and multi-llm-provider-go, and React Flow for the workflow canvas.