multi-agent · production-ready · repo-aware

The AI coding agent that
ships production code,
not textbook answers

Elith reads your prompt, understands the intent, spins up the right swarm of specialized agents — devops, dev, qa — and orchestrates them until the codebase is genuinely production-ready. Every model. Any repo. One framework.

~/projects/payments-api — elithzsh

$ elith "harden checkout for black friday traffic"

→ intent parsed: scale + reliability

→ spawning agents:

[devops] capacity & infra review

[dev] refactor hot path, add idempotency

[qa] load tests + chaos scenarios

→ context engine: loaded 7 files from 412

→ proposing architecture: CQRS read replicas + queue backpressure

✓ pipeline ready: devops → dev → qa → devops → prod

// preview

See it in action

The Elith TUI — agents, runtime diagnostics, and context loaded into a single keyboard-driven workspace.

~/.elith — elithzsh
Elith CLI session showing multi-agent orchestration, runtime diagnostics, and recommended actions

// the problem

§ 01 — why ai-assisted dev still fails
01

Textbook answers

Ask any LLM to architect a system — you get a generic MVC/REST diagram. Never a hybrid CQRS because your read:write is 20:1. Never a strangler fig because you can't stop the business.

02

Repo-blind models

Claude, Gemini, GPT are powerful reasoners — but they can't read your files, run your tests, or commit changes. You paste, they guess, you copy back. Context bloats. Quality drops.

03

Single-agent ceilings

One model trying to do dev, qa and devops simultaneously will mediocre at all three. Real shipping requires specialists handing off work — not one generalist juggling roles.

// the pipeline

§ 02 — prompt to production, no humans in the loop

Elith doesn't generate code — it orchestrates a team. Each agent owns a phase of delivery. They hand off, review each other's output, and loop until the work actually meets a production bar.


   ┌──────────┐   ┌──────────┐   ┌──────────┐   ┌──────────┐
   │  PROMPT  │──▶│  INTENT  │──▶│  AGENTS  │──▶│  ROUTER  │
   │  parser  │   │ analysis │   │  spawn   │   │ pipeline │
   └──────────┘   └──────────┘   └────┬─────┘   └────┬─────┘
                                      │              │
        ┌─────────────────────────────┴──────────────┘
        ▼            ▼            ▼            ▼
   ┌─────────┐  ┌─────────┐  ┌─────────┐  ┌─────────┐
   │ DEVOPS  │◀▶│   DEV   │◀▶│   QA    │◀▶│  PROD   │
   │ scaffold│  │ implement│ │ verify  │  │ deploy  │
   └─────────┘  └─────────┘  └─────────┘  └─────────┘
                     ▲            │
                     └── loop ◀───┘   (qa fails → back to dev)
phase 1
Intent
Parse the prompt. Map it to a delivery shape: feature, refactor, harden, migrate.
phase 2
Scope
Context engine loads 4–6 relevant files (not all 312). Builds a focused task packet.
phase 3
Execute
Agents work in parallel. Devops prepares infra, dev writes code, qa drafts tests.
phase 4
Verify
QA runs the suite. Failures route back to dev. Loop until green, then promote.

// the swarm

§ 03 — specialists, not generalists
agent~/devops

[Devops]

  • scaffold infra
  • ci/cd pipelines
  • secrets & env
  • containerize
agent~/architect

[Architect]

  • read full repo
  • propose 2-3 designs
  • tradeoff analysis
  • migration path
agent~/dev

[Dev]

  • implement modules
  • follow repo patterns
  • wire integrations
  • open PRs
agent~/qa

[QA]

  • generate tests
  • run suite
  • chaos & load
  • report regressions
agent~/reviewer

[Reviewer]

  • read diff
  • flag risk
  • enforce standards
  • summarize change
agent~/docs

[Docs]

  • update readme
  • write adrs
  • api references
  • changelogs

// every model

§ 04 — bring your own brain

Pick the model you trust. Elith's skill layer gives each one the same repo awareness — read files, search code, run tests, commit changes — through agent tool calling. No paste-and-pray.

model
provider
interface
Claude
Anthropic
tool calling
GPT-4 / Codex
OpenAI
tool calling
Gemini
Google
function calling
Mistral
Mistral
skill layer
Local LLM
Ollama
skill layer
Any OpenAI-compat
Custom endpoint
skill layer

// textbook vs elith

§ 06 — same question, different answers
prompt
standard llm
elith
"Structure this service"
Generic MVC/REST
Hybrid CQRS — your read:write is 20:1
"Migrate legacy system"
Full rewrite
Strangler Fig — you can't stop the business
"Handle state"
Redux, default
Event sourcing — you need audit trails
"Add caching"
Redis everywhere
Per-route TTL + stale-while-revalidate

// install

§ 05 — one command, any package manager
$ curl -fsSL https://elith.silohq.tech/install.sh | sh
auto-start
Backend boots & survives CLI exit
cross-platform
macOS · Linux · WSL
service control
elith service {start|stop|status}