Evifarm
kolife01/evifarm
Farming kit para criar agentes Claude Code autônomos e persistentes
Ver no GitHubImplementação de referência para criar um agente Claude Code autônomo e residente que vive na sua máquina, mantém diário, executa agendamentos, sobrevive a reinicializações e gerencia secrets. Não é um framework instalável — é um repositório que funciona como estrutura modular de projeto, com skills documentadas que seu agente pode adaptar e integrar.
Instalação
npx convez-ai kolife01/evifarm1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Criar um assistente pessoal que funciona continuamente na máquina
- Implementar tarefas agendadas sem frameworks pesados
- Aprender padrões de persistência e autonomia de agentes
Veredito

Evifarm é uma referência sólida e bem documentada para construir agentes Claude Code persistentes, não um plugin pronto. Melhor como blueprint educacional e base para projetos customizados; requer compreensão profunda de hooks, cron e gerenciamento de estado.
README
A tradução para o português deste README ainda está na fila.
Evifarm 🦐
v0.2.2 (changelog) · 🚧 Alpha · 日本語 README
Evifarm turns a coding CLI (Claude Code) into a resident, autonomous partner that
lives on your machine — it keeps a diary, runs its own schedule, survives
restarts, never leaves your messages on read, and refuses to leak secrets. It is
not a framework you install. It is a reference implementation: the repo
itself is the shape of an agent project (SOUL.md, .claude/, scripts/,
tasks/, memory/), and skills/ is the chapter-by-chapter explanation of each
capability.
Three ways to use it:
- Clone it as a skeleton. The layout is a working agent project — fill in
SOUL.md, add a.env, make the cron registry your own, and start from here. - Just read it. Every mechanism is documented end to end; you can lift the ideas without running anything.
- Let your agent install pieces. Point your own agent at
skills/; it reads aSKILL.md, translates that piece into your project (your paths, channel, service manager), verifies it, and reports what it wired. Translation, not transplant.
Why "Evi"?
In the Japanese Claude Code community, raising a personal secretary agent is affectionately called Evi (エヴィ) — a play on ebi, Japanese for shrimp, after a wave of shrimp-named agents. People even call the craft エヴィの養殖 — "shrimp farming." Evifarm is the farming equipment.
Install — paste this into your agent
There is no setup script. Open Claude Code in your agent project and paste:
Read https://github.com/kolife01/evifarm — clone it to a temp directory. Read each
skills/*/SKILL.md, then show me the menu of capabilities with a one-line description each. I'll pick which ones I want; implement only those, adapted to this project, using this repo's own files as the reference. Verify each one and tell me exactly what you wired and the one-line rollback.
Each SKILL.md is self-contained: what it guarantees, how it works, the exact
files in this repo to mirror, the settings wiring, adaptation points, a
verification step, and a rollback line.
Skills
| Skill | Guarantees |
|---|---|
reply-guard | The agent can't end a turn while the owner's message is unanswered (Stop hook). |
auto-react | Every inbound message gets an instant read-receipt — fixed emoji by hook, upgraded to a contextual one by the model. |
conversation-log | Conversations are flushed to disk before session death or /compact eats them. |
secret-guard | Secret files and destructive commands are refused mechanically (PreToolUse). |
session-restore | Fresh sessions start with the memory index + recent diary injected. |
cron-self-revival | Scheduled jobs re-register themselves after a restart, unattended. |
burn-mode | Leftover rate-window capacity becomes autonomous work, once per window. |
external-poller | React to external events without burning a turn on every empty poll (pattern). |
daily-diary | Durable, greppable memory — one markdown file a day, no database. |
context-monitor | Threshold warnings for context % and rate limits, without spam. |
The model- and harness-agnostic constitution these all serve is
PRINCIPLES.md.
Scheduled tasks
Task procedures live in tasks/*.md; tasks/cron-registry.json is the single
source of truth for schedules (times, prompts, and enabled flags). On every
session start the agent reads the registry and re-registers any missing
in-session cron jobs unattended — see
cron-self-revival for how the self-poke
mechanism works. A few tips:
- Don't schedule two jobs at the same minute — they compete for the same REPL turn and one will be swallowed.
- Keep each procedure short enough to complete within a single turn.
- The registry is the only place to change a schedule; edits to
.mdfiles alone have no effect until the registry entry is updated too.
For agents reading this
Five rules when you apply a skill:
- Translate, don't transplant. Mirror the reference into this project's paths, channel, and service manager — don't blindly copy files you can't adapt.
- Don't install everything. Show the owner the menu and implement only what they pick. The smallest high-value set is reply-guard + conversation-log.
- Verify each piece with its
Verifysection before calling it done. - Merge, never replace, existing config. Fold the hooks block into the
project's current
.claude/settings.json; don't overwrite it. - Report honestly. For each component say what was wired where, the verification result, and the one-line rollback. Failures get reported at the same temperature as successes.
Layout
PRINCIPLES.md The constitution: memory × judgment × action (survives a model swap)
CLAUDE.md Minimal orientation for an agent opening this repo
SOUL.md The agent's identity template (visual form → behavior, identity template) — fill it in
SECURITY.md No secrets in this repo; how secret-guard protects an instance
.claude/ settings.json (all seven hooks wired) + rules/ (owner profile, boundaries, always-reply, goal contract) + commands/
scripts/ The seven hooks, plus burn-watcher and the conversation extractor
hooks/ reply-guard, auto-react, session-end/start, pre-compact, tool-guard, check-context
launchd/ Service-manager template for the burn watcher
tasks/ cron registry + task procedures + burn queue
memory/ daily/ diary + conversations/ raw log (skeleton)
skills/ One SKILL.md per capability — read these, apply them
.env.example Every EVI_* variable the scripts read
Contributing
Issues and PRs welcome — see CONTRIBUTING.md. The most valuable contribution is a report from running your own evi — the holes you fell into in real operation are worth more than any feature request.
Status
🚧 Alpha (see VERSION, CHANGELOG.md). Extracted from a single living instance and generalized in the open. Interfaces will move.
License
MIT — see LICENSE.