convez.ai
AgenteAutomaçãoAvançado
richardtkemp

Foci

richardtkemp/foci

Agentes IA leves em Go com isolamento de secrets e arquitetura cache-first

Ver no GitHub
BomPluginScore · de 0 a 100

Platform de agentes IA em Go, inspirada em OpenClaw, otimizada para eficiência de recursos e segurança. Oferece isolamento de secrets em nível de SO, cache-first architecture, piping de ferramentas e suporte nativo a Telegram com footprint mínimo (~35 MB).

Instalação

$npx convez-ai richardtkemp/foci

1 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

Estrelas
6
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 dia
Versão
—
Issues abertas
0
Contribuidores
3

Casos de uso

  • Orquestração de agentes de coding em Telegram
  • Automação com controle de segurança rigoroso de credenciais
  • Aplicações com restrições severas de memória e latência

Veredito

Foci é uma alternativa técnica ao OpenClaw para quem prioriza eficiência, auditabilidade e segurança sobre amplitude de features. A arquitetura cache-aware e isolamento de secrets em SO são diferenciais legítimos, mas a maturidade e ecossistema ainda são inferiores.

README

A tradução para o português deste README ainda está na fila.

Foci

One binary. ~35 MB idle. No framework.

AI agents on Telegram, written in Go.

Quick Start · Design · Docs


Inspired by OpenClaw, built from scratch in Go — single binary, cache-first, with OS-level secret isolation. Session branching, tool piping, and coding agent orchestration go well beyond the original. Built for Anthropic, but any OpenAI-compatible endpoint works.

Quick Start

git clone https://github.com/richardtkemp/foci.git && cd foci && ./setup.sh

See docs/INSTALL.md for prerequisites, options, and next steps.

Background

OpenClaw is the established, full-featured choice in this space — broad provider support, native apps on every platform, a marketplace of 13,000+ skills, and a large community. Foci makes different bets. Where OpenClaw optimizes for breadth, foci optimizes for depth: cache-aware prompt architecture, OS-level secret isolation, and a codebase small enough that one person can audit the whole thing. See docs/COMPARISON.md for a detailed feature comparison.

OpenClawFoci
RuntimeNode.js + TypeScriptGo, single binary
Memory~500MB+ idle~35 MB
Dependencies~560 packages (594MB)15 direct modules
StartupSeconds (transpile + boot)Instant
ConfigYAML + env + scattered filesOne TOML file
Cache strategyNot cache-awareDay-zero architectural

Fewer moving parts means fewer surprises. Secret management follows the same principle: OS-level isolation, domain-locked credentials, redaction at every layer — protection that doesn't depend on trusting the model.

Design

Cache-first architecture — Built around Anthropic's prompt cache. Character files form a stable prefix, session branching shares cached context, and the system actively avoids invalidation. More work per dollar. → docs

OS-level secret isolation — Secrets are readable only by a dedicated group. Child processes have that group dropped. Domain-locked, output-redacted, with optional Bitwarden vault integration. → docs

Tool result guard — Large tool outputs are truncated before entering context, with full results saved to disk. Cache stays intact, context window stays clean.

Tool piping — Tools are exposed as shell functions, composable with Unix pipes. foci_web_search "latest golang release" | foci_spawn "summarize" --model haiku | foci_send_to_chat — three tools, one shell call, no intermediate data in context. → docs

Facet — /facet forks your session to a second Telegram bot — same agent, same context, parallel thread. Both share the cached prefix, so the fork is cheap. → docs

Multi-agent, single process — Multiple agents share one binary with separate workspaces, identities, and Telegram bots. No containers, no orchestration. One TOML file.

Coding agent orchestration — Tmux management and coding agent control are structured tool calls. Start sessions, send instructions, watch for inactivity, read output.

Compaction that preserves personality — Context compression keeps goals, reasoning, corrections, emotional tone, and technical state — not just a generic summary. Configurable per-agent. → docs

Mid-turn nudges — Behavioral reminders extracted from character files by the LLM and injected mid-conversation. Five trigger types keep the agent on-character without busting the cache. → docs

Memory out of the box — Daily markdown files + curated MEMORY.md. No vector DB, no embeddings. The agent reads, writes, and prunes its own memory. → docs

Who foci is for

  • You run agents in production and care about cost per conversation
  • You want to audit exactly what your agent can access — and prove secrets never reach the model
  • You prefer a small, focused system with sensible defaults over a comprehensive one you have to configure
  • You're comfortable with Telegram or API-level interaction

Who should use OpenClaw instead

  • You need native apps on macOS, iOS, or Android
  • You want a built-in skill marketplace for discovery and install (foci can run OpenClaw skills — they're text files — but has no marketplace)
  • You need broad provider failover across many LLM backends
  • You want a large community and established support ecosystem

OpenClaw is good software. If those are your priorities, use it.

Requirements

Go 1.24+Build from source
Telegram bot tokenMessage transportCreate via @BotFather
Claude CodeSubscription tracking + coding agent orchestrationdocs/AUTH.md
bashset -o pipefail, tool-piping shell functionsFalls back to sh
Groq API keyVoice input (Whisper STT)Free tier
Brave Search API keyweb_search toolFree tier
edge-tts + ffmpegVoice output (TTS)pip install edge-tts
Bitwarden CLIDynamic secret managementdocs/BITWARDEN.md

See docs/DEPENDENCIES.md for full details.

Documentation

  • SPEC.md — Full specification (source of truth)
  • WIRING.md — Internal architecture and wiring
  • CACHING.md — Cache architecture and preservation
  • COMPARISON.md — Feature comparison with OpenClaw and Nanobot

Stats

~67k lines of Go (~158k with tests) · 1,670+ commits · 15 dependencies · 47 packages

License

PolyForm Noncommercial 1.0.0 — free for personal, academic, and non-commercial use. Commercial use requires a separate license. Contact Richard Kemp for details.

Ver o original no GitHub

Bom para

Relacionados