cd /news/ai-agents/muxlio-durable-agent-workflows · home › topics › ai-agents › article
[ARTICLE · art-172863] src=github.com ↗ pub= topic=ai-agents verified=true sentiment=↑ positive

Muxlio – Durable Agent Workflows

Muxlio released an Apache-2.0-licensed tool that runs multi-step workflows across AI agents and command tools, saving progress so a client can disconnect while the coordinator keeps running. The tool installs via Homebrew on macOS or Linux and bundles a CLI, MCP server, adapters and web UI; its example "implement-features" recipe runs Implement → Review → Run checks → Commit locally → Next feature, pausing for attention after ten unsuccessful verification attempts. Muxlio's author says he uses it to develop Muxlio itself, running it on a Mac mini and queuing work through an assistant.

key takeaway

Muxlio's implement-features recipe runs build and tests before each local commit and pauses after ten failed verification attempts, letting a developer review changes and test output before pushing.

read4 min views13 publishedNov 19, 2026
Muxlio – Durable Agent Workflows
Image: Michielbdejong (auto-discovered)

Build a workflow. Let your agents run it.

Documentation · Get started · Implement-features recipe · Apache-2.0

Muxlio runs multi-step workflows across AI agents and command tools. You or your assistant define the steps, their inputs, and the checks they must pass. Muxlio executes the workflow, saves progress, and gives you a place to follow the work.

Use it for work that needs several steps, checks, or hand-offs. Your client can disconnect while the coordinator continues running. You can return to the run's status, logs, outputs, and any failure that needs attention.

I use Muxlio to develop Muxlio itself. The implement-features recipe works through features one at a time:

Implement → Review → Run checks → Commit locally → Next feature
                         │
                         └─ if checks fail: Fix → Run checks again

The workflow runs the repository's build and tests before committing. After ten unsuccessful verification attempts it s for attention. You can inspect the change, review notes, test output, and execution history before deciding to push. Each verified commit becomes the starting point for the next feature.

To run the batch in a private Git clone, follow the Docker Sandbox walkthrough. One sandbox worker offers Claude Code, Codex, Cursor, and shell commands. The example takes two dependent features from setup through preflight, execution, and fetching the commits. The live recipe also lets you add features to the same branch while the batch is running. If you set up both workspaces, choose localhost to run on the host or sandbox to run on the sbx worker.

I keep Muxlio running on a Mac mini, queue work through my assistant, and review progress later. You choose the machine: your workstation, another computer, or a server. The execution host needs to stay running; your chat session does not.

This is one workflow, not a special product command. Adapt the recipe's tasks and checks to your own project, or build a workflow for research, data processing, documents, or other work.

With Homebrew available on macOS or Linux:

brew tap Muxlio/muxlio https://github.com/Muxlio/muxlio
brew install --HEAD Muxlio/muxlio/muxlio
muxlio version

This builds the current main branch and installs the CLI, MCP server, adapters, and web UI. Homebrew handles the build dependencies; you do not clone the repo. Install and sign in to the agent CLIs your workflows will use separately.

With Muxlio installed, initialize the workspace where work should happen:

muxlio init --client kiro --scope project
muxlio service start

Choose --client for your authoring assistant. init allows all bundled executors by default, including agent tools, commands, workspace writes, and network access. Change executors in ~/.muxlio/config.json to restrict permissions; repeated init preserves explicit policies and disabled executors. init prepares the workspace settings, credentials, agent instructions, and MCP settings. The service starts the coordinator, local worker, UI, and configured MCP listener together.

Connect your assistant through MCP, then ask it to build a workflow for your task. Describe the inputs, the steps to run, and the output you need. Have it validate the workflow and show you the plan before starting the run. Developers can also author workflows directly with the TypeScript SDK.

The installed service does not need Muxlio's source checkout. Workers need the files, authenticated agent CLIs, and tools required by your workflow. Source-build instructions are available if you prefer to build the runtime yourself.

Open http://localhost:7600 or ask your assistant for run status. Inspect the execution graph, attempts, live logs, and generated artifacts. , resume, or retry when appropriate. Uncertain external effects require reconciliation before work can safely repeat.

To launch a registered workflow from the dashboard, select its version and an authorized workspace, enter the inputs, and run preflight. See the dashboard guide.

muxlio service stop
muxlio service restart
Start here Go deeper
Which adapter should I use? — native CLIs, ACP, and A2A Adapter capability reference
Service Guide — setup and service controls Service reference
Docker Sandbox walkthrough — run dependent features in a private clone Worker protocol
MCP Integration — build and run from your assistant MCP tools and protocol
Observability UI — follow progress and results Operations API
Benchmarks — what the suite checks Benchmark methodology

Explore the architecture, execution and security contracts, and CLI, SDK, and configuration reference. The example recipe and authoring guide show how workflows are put together.

muxlio init --client NAME --scope project --workspace /absolute/project installs the discoverable Muxlio skill and its references in that project. Use --scope user for discovery across projects. See client skill setup and invocation for supported paths, Cursor version handling, and preservation of existing edits.

Installed releases support offline JSON authoring: run muxlio template echo > workflow.json, then muxlio check workflow.json. The echo template needs no executor; muxlio template task shows a generic task with input and output bindings. Templates are embedded in the binaries and shipped in share/muxlio/templates. With MCP, use workflow_template, then check_plan and start_run. No Node packages or registry access are needed for JSON authoring, validation, fake testing, or submission.

See the installed authoring guide for an empty-workspace example and submission instructions. Optional TypeScript/ JavaScript compilation executes arbitrary code, requires project-local dependencies installed from your pinned lockfile, and never implicitly downloads packages.

── more in #ai-agents 4 stories · sorted by recency
── more on @muxlio 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/muxlio-durable-agent…] indexed:0 read:4min 2026-11-19 · —