Take your agents live. Turn Claude Code, GitHub Copilot, and OpenAI Codex agents into scheduled and file-triggered local automations, without moving them to another agent platform.
Definitions can be conforming Agent Skill directories or flat Markdown files in configured repository directories. Agents Live reads their namespaced execution metadata, adds local triggers, and repairs drift using standard host tools.
---
name: markdown-polisher
description: Polish Markdown documents when they change.
metadata:
agents-live.schema-version: "2"
agents-live.selector: "claude"
agents-live.mode: "write"
agents-live.watch: "docs/** debounce 1s"
---
Correct spelling, grammar, and Markdown formatting errors in the selected files.
Preserve their meaning, links, code, and frontmatter. When a `Files changed:`
list is present, process only those files.See Installation for required host tools and platform-specific instructions.
curl -fsSL \
https://github.com/johnshew/agents-live/releases/latest/download/install.sh | sh
agents-live init
agents-live start markdown-polisherThe watcher sleeps until a file changes, then runs the agent immediately with
the changed paths. Add or edit a Markdown file under docs/, then open the
file to see the fixes.
Manage the running agent with status and stop:
agents-live status
agents-live stop markdown-polisherThere is no polling interval or clock tick. The agent runs only when the operating system reports a change in the watched directory.
There is no listener service, separate application runtime, or database to
deploy and maintain. The core stack is the Claude Code, GitHub Copilot, or
OpenAI Codex CLI you already use, uv, and your host scheduler and file-watch
facility.
Cron-only agents have no persistent process. A file-watch agent uses one small local watcher. There are no externally reachable ports or databases. Custom post-processors and plugins may bring their own dependencies; Agents Live core does not require them.
Execution modes make write access explicit:
planis read-only. The agent emits JSON for a validated post-processor to apply.pipelinelimits the agent to a schema-checked data channel shared with your pre-processors and post-processors.writegrants full write access as an explicit per-agent choice.
This is tool policy, not a sandbox. Agents still inherit the permissions of your local account and agent CLI.
The example uses write so it can fix documents directly. For tighter
control, use the complete plan and pipeline Markdown-polisher
examples. They apply the
same correction task through validated, deterministic write boundaries.
Agents Live supports Linux, WSL, and native Windows. macOS is currently untested. First install and sign in to at least one supported provider CLI:
On Debian, Ubuntu, or WSL, install the host tools used by schedules and file watchers, then run the latest stable installer:
sudo apt-get update
sudo apt-get install -y cron inotify-tools
curl -fsSL \
https://github.com/johnshew/agents-live/releases/latest/download/install.sh | sh
agents-live --repo /path/to/repository initcron runs scheduled agents and automatic maintenance. inotifywait is only
needed for file or directory watches. The installer links agents-live and
al from ~/.local/bin to the stable current commands and adds that
directory to supported shell profiles. Open a new shell if the current one did
not already include ~/.local/bin on PATH.
The installer refuses to replace an existing ~/.local/bin/agents-live or
~/.local/bin/al that does not point to this installation. Remove or rename
the conflicting command, then run the installer again.
On WSL, use these Linux instructions inside the distribution. The first convergence also stages and verifies Windows-side liveness so scheduled work can wake the distribution without an open terminal.
Run PowerShell as your normal user. Install a provider CLI if one is not already available:
winget install Anthropic.ClaudeCode
# Or: winget install GitHub.CopilotRun the latest stable installer, then initialize a repository:
irm https://github.com/johnshew/agents-live/releases/latest/download/install.ps1 | iex
agents-live --repo C:\path\to\repository initThe installer updates both the current PowerShell process and the persistent
user PATH. Native Windows uses Task Scheduler and directory change
notifications, so no separate scheduler or watcher package is required.
agents-live upgrade
agents-live uninstallActivation does not wait for in-flight agent runs. They finish on their original immutable version while new launches use the selected version; shared per-agent locks prevent overlap. Version removal and collection preserve versions still in use. Runtime maintenance and conflicting mutations can still refuse activation. Watcher-triggered runs also finish unharmed: busy watcher trees are preserved, and replacement of their subscription is deferred until completion and handoff. Convergence rechecks ownership before spawning and removes idle duplicate watcher trees, preferring the selected generation.
Source plugins load directly from their declaring repositories. Identical
same-name copies across registered repositories share one imported module;
conflicting plugin or provider names report the repositories and plugin files.
Use agents-live doctor --all-repos to inspect each declaration's health.
See the command reference for pinning, rolling back, or removing installed versions. The diagnostics guide covers installation conflicts, existing repositories, and repair.
Run diagnostics retain lock acquire waits and hold times, planned versus actual
clock launch times, and the producing runtime version and generation.
Use agents-live logs --sql for max/p95 timings and the clock_fires view for
missed or skipped fires within recorded schedule intent. The dashboard's Timing
panel exposes the same queries plus watcher exit codes, bounded stderr tails
and termination reasons. Maintenance checks actual watch-loop progress rather
than only process markers. See the diagnostics guide.
Definitions live under a registered repository's Agents/ directory by
default, and Agents Live also searches .claude/skills/, .github/skills/,
and .agents/skills/, claiming a skill there only when it carries
agents-live. execution metadata. Set agent_directories = ["foo"] in
.agents-live.toml to also discover immediate foo/<name>.md files and
foo/<name>/SKILL.md bundles.
Register another repository with agents-live init --repo <path>. Once
registered, run, start, stop, and status fall back to the other
registered repositories when a name is not present locally.
Cross-machine assignment is optional. Repository registration and ownership
backend installation leave a project local-only. Run
agents-live ownership enable to validate the backend and owners document
before enabling transfers; agents-live ownership status reports the mode.
See the command reference for repository workflows, health checks and repair, upgrades, dashboards, shell completion, plugins, ownership, and multi-repository operations. The architecture guide covers runtime, safety, persistence, and maintenance behavior.
Every workflow is an ordinary CLI command.
Design documents and the high-level backlog for the project itself live in docs/; they are not installed with the skill.
Bug reports and pull requests are welcome in Issues.