Self-hosted server that runs the automated editorial systems, Ciencia a Mordiscos among them: Go agents on systemd timers, an in-house LLM gateway, reviewers with veto power, and human approval from Linear. CI/CD with SAST, severity-routed monitoring alerts, verified encrypted backups, and decisions recorded as ADRs.
Go agents in production
20+Go agents in production
lines of Go
75k+lines of Go
Built with Claude Code
Trellium's code, tests, and ADRs are written in Claude Code sessions.
The pieces
Pieces built inside Trellium. Dev-Agent has since been retired, and Del Champi is on hold.
Self-Hosted Automation Lab
Self-hosted server running three editorial systems in production (video and WhatsApp audio newsletters) and a generative music channel on hold, on a 75k+ LOC Go monorepo (plus another 75k+ of tests) with 40+ binaries, 45+ systemd timers, and an in-house LLM gateway. 15+ containers with least-privilege hardening behind a reverse proxy with TLS and HSTS, automatic threat detection and blocking, CI/CD with SAST, CVE scan, and E2E smoke tests, custom monitoring with 25+ checks and severity-routed alerts, encrypted off-site backups verified weekly, and continuous auditing of code, dependencies, and Cloudflare configuration against a baseline. Decisions recorded as ADRs.
AI Agent Orchestration Platform
Go binaries that operate the server: health (25+ checks across 9 families), updates (images, GitHub, npm, and system packages), code auditing, and Cloudflare configuration against a baseline. Circuit breakers, per-file mutex, retry policies, and severity-routed alerts (push or Linear issue). systemd timers, the critical ones with OnFailure=, and E2E smoke tests when a deploy touches their dependencies.
Claude API Gateway
In-house LLM gateway in Go: priority queues, SSE streaming, rate limiting, per-service authentication, and token/cost accounting per service and model in SQLite. Runner abstraction that lets agents run through Claude CLI or the Go SDK (prompt caching, Batch API) without changing the call site; production runs on the CLI.
Dev-Agent: FSM Multi-Agente para Issues Linear
System that picked up Linear issues labeled dev-agent and processed them through 6 phases (Planner Opus → Builder Sonnet → Auditor Opus → Fixer → Re-auditor → Tester) driven by Linear labels. It opened GitHub draft PRs with human gates at plan and integration; the Auditor waited for CI and classified findings CRITICAL/HIGH/MEDIUM/LOW. Retired in June 2026.
Del Champi — Generative Electronic Music
In-house generative electronic music engine: 3-hour seed-based sessions with direct DSP (no MIDI or soundfonts), act-based narrative arc, and per-genre generative visuals. Quality measured with CLAP, FAD, and fatigue proxies, plus logged blind listening; gates that ran in CI. Episodes published on YouTube; channel on hold.