ClaudeMap

agent-purge

One persona to govern multiple coding agents: a local-first prompt-configuration tool giving Codex, Claude Code, OpenCode, pi, DeepSeek Harness, WorkBuddy, Kilo Code and others a unified persona gateway — web console plus CLI, four injection modes, traceable and reversible changes (MIT)

Plugins & Agentspersona managementprompt configurationcoding agentsweb consolelocal first

agent-purge is ZiChuanLan's local prompt-configuration and persona-orchestration tool (1,563 stars as of writing, 2026-11-11): personas live as Markdown files in ~/.agent-purge/personas/ on your machine, the body becomes the prompt, and the CLI and web console (default 127.0.0.1:7788) share one persona store and state. Four injection modes, settable per harness: replace (default — replaces the host identity line where readable), strict (replaces the host text in controllable entries), append (keeps host instructions and adds the persona), off (stops injection); hot-switch channels reflect changes on the next turn, static entry points need a new session. Four application-scope channels: install (writes managed host config/plugin), shim (forwarding through a same-named PATH command), slash (an in-session /purge extension), project (per-directory rule files). Every managed write is ledgered with its restore basis, and recovery conflicts fail loudly instead of overwriting. The English README describes the audience as 'coding agents such as Codex, Claude Code, OpenCode, pi, DeepSeek Harness, WorkBuddy, and Kilo Code', and the support table in the canonical Chinese README lists per-harness verified versions with the explicit warning that 'configuration written, historical tests passed, and actually taking effect in the current session are different states.'

What it does

  • Local-first with zero production dependencies: personas as editable Markdown in ~/.agent-purge/personas/; upgrades never overwrite the data directory
  • Four injection modes (replace / strict / append / off), settable globally or per harness (agent-purge use <id> --for codex)
  • Four channels: install (managed host config/plugin), shim (PATH forwarding), slash (/purge in-session), project (per-directory rules) — arm a harness with agent-purge arm <h>
  • Traceable and reversible: managed writes and stashed backups go into a ledger; recovery conflicts report explicitly
  • Per-harness verification table with pinned versions: pi/omp and OpenCode hot-switch verified; Claude Code / Codex launch entry verified, hot-switch unverified
  • An optional magpie gateway plugin lives in magpie-plugin/, not the repo root

When to use it

For developers who move between several coding agents but want one collaboration persona everywhere: edit once, apply everywhere; also for per-directory project rules, one-shot launches with a specific persona (agent-purge codex --mode append), and teams distributing a shared convention across harnesses.

Getting started

# requires Node.js 20+ and the coding agents you want to manage installed locally
npm install -g agent-purge
agent-purge web

# or without a global install:
npx agent-purge web

# the console listens on 127.0.0.1:7788 by default; edit the neutral starter persona in the persona library,
# pick persona and mode in the console, then enable the harnesses you want to manage

# everyday commands:
agent-purge ls                  # list personas
agent-purge use <id>            # switch and sync the selected scope
agent-purge mode append         # change the global mode
agent-purge use <id> --for codex  # assign a persona to Codex only
agent-purge status              # see what's armed

Things to know

The README's own compliance section is extensive and worth reading in full: it is a local multi-host configuration-management and debugging tool; it ships no jailbreak or security-evasion prompt payloads; persona injection can change model behavior, so back up configs and test in isolation before important use; and an enabled persona is sent to the harness's configured model service with host requests — as the README puts it, local configuration does not mean offline AI execution. Platform status: regressions were run on Windows and WSL/Linux, macOS is not yet machine-verified; Hermes and zcode are not counted as passed this round. Request-level verification covers only the pinned versions in the table (Claude Code 2.1.316 / Codex 0.166.0: launch entry verified, hot-switch unverified). strict does not mean the model receives only the persona: the actual role may be system, developer or project context. The web console listens on 127.0.0.1 only — do not expose it to the public internet. When a restore conflict appears, keep the files and credentials the error names and compare before deleting backups. MIT. 1,563 stars as of writing (2026-11-11).

Repository

ZiChuanLan/agent-purge

Charted

Related on the map

  • clawd-conduit

    Plugin / Agent

    Wire all 15 Claude Code hook events into the Clawd on Desk desktop pet — and turn permission requests into a blocking desktop card.

    • hooks
    • desktop
    • notifications
    • claude-code
  • CometixCode

    Plugin / Agent

    An unofficial Rust 1:1 reimplementation of the Claude Code terminal UI, ported file-for-file onto an iocraft TUI framework.

    • terminal-ui
    • rust
    • reimplementation
    • unofficial
  • Abide

    Plugin / Agent

    Hook-level enforcement for the natural-language rules in AGENTS.md and CLAUDE.md — every edit is judged and violations are repaired in the same turn.

    • hooks
    • code-review
    • rules
    • agents-md