# AgentSwitch

> ax is a daemonless switchboard for CLI coding agents. One picker, one verb set, local and remote sessions. No daemon, no server, no runtime of its own.

AgentSwitch ships as a single Go binary called `ax`. It launches, watches, detaches, reattaches, and coordinates coding agent sessions from one terminal picker, and gives every session a set of shell verbs so workflows compose out of plain shell. The pages below are also served as clean markdown: request the `.md` twin directly, or send `Accept: text/markdown` to any HTML page.

## Docs

- [Quickstart](https://agentswitch.org/index.md): install ax per OS, set a default harness, run your first session, and launch a project coordinator.
- [Manual](https://agentswitch.org/manual.md): complete reference for the picker, control-plane verbs, fences, remote hosts, running agents safely, and configuration.
- [Tutorials](https://agentswitch.org/tutorials.md): worked runs with captured terminal output, from a first session to a coordinated multi-agent task.
- [Recipes](https://agentswitch.org/recipes.md): copy-paste playbooks that run as written, plain shell plus ax.
- [FAQ](https://agentswitch.org/faq.md): common questions, what ax is for, how it differs from an agent harness, cost, composition, and which harnesses it supports.

## Recipes

- [Cost-tiered model routing with escalation](https://agentswitch.org/recipes.md#route-a-task-up-through-model-tiers): run on the cheapest tier first, escalate on a failed accept check, cap spend with token fences.
- [Cross-agent blackboard](https://agentswitch.org/recipes.md#share-state-between-agents-with-a-blackboard): two or more workers coordinate through a shared JSON file on disk, no queue server.
- [Scattered sources to queryable vault](https://agentswitch.org/recipes.md#fan-out-workers-over-a-file-collection): fan N workers in parallel over a corpus into a shared vault, then verify completeness.
- [Behavior audit](https://agentswitch.org/recipes.md#audit-session-history-for-behavior-patterns): read recent transcripts, extract recurring patterns, produce a human-reviewable lessons document.
- [Email triage with sentinel gate](https://agentswitch.org/recipes.md#triage-email-act-on-everything-notify-on-exceptions): act on every message in a maildir and notify only when a human is needed.
- [Parallel design pipeline with human-gated publish](https://agentswitch.org/recipes.md#draft-listings-and-hold-for-human-approval-before-publishing): fan out one worker per niche, then pause at `ax ask` until a human approves.
- [Deterministic pre-script with LLM formatter](https://agentswitch.org/recipes.md#collect-data-in-a-script-format-it-with-an-agent): a script collects ground truth, the LLM only formats it, a sentinel gate decides delivery.
- [Scheduled chain](https://agentswitch.org/recipes.md#schedule-any-recurring-workflow): a cron trigger runs an ax worker, checks the output for a sentinel word, and delivers only when actionable.
- [Produce-judge-iterate loop against a rubric](https://agentswitch.org/recipes.md#run-a-produce-judge-iterate-loop-against-a-rubric): a producer emits a deliverable, a reviewer scores it against a rubric, violations fold into the next iteration, loop stops on pass or cap.

## Optional

- [Full documentation](https://agentswitch.org/llms-full.txt): every page concatenated as one markdown file.
- [Source repository](https://github.com/agentswitch-org/ax): the ax binary, recipes, and multiplexer bindings.
