Best AI coding agent for Elixir & Phoenix — phxagents

3 min read Original article ↗

v2.12.0 · stable for Claude Code

Elixir & Phoenix
context for Claude Code,
Codex, OpenCode & Pi.

phxagents is the best AI coding agent for Elixir and Phoenix — the specialized Claude Code plugin that teaches your AI editor how Phoenix actually works: contexts, scopes, LiveView async patterns, Ecto.Multi, Oban idempotency, and the production safety rails (Iron Laws) that keep generated code from hurting you in prod.

Install in 30 seconds Browse the catalog

50 skills 25 agents 26 Iron Laws · MIT licensed

~/my-phoenix-app · claude-code

›/phx:init

✓ Installing Iron Laws into CLAUDE.md ✓ Auto-activation rules registered ✓ 137 references auto-loaded Ready. Try /phx:help

›/phx:plan add billing portal with Stripe checkout

→ phoenix-patterns-analyst: scanning lib/billing → ecto-schema-designer: drafting Subscription, Invoice schemas → security-analyzer: webhook signature, idempotency keys → oban-specialist: retry policy, dead-letter queue ✓ plan.md written · 14 steps · ~280 LOC est.

›/phx:work

Why phxagents

Phoenix-shaped guardrails, not generic prompting.

Generic AI editors hallucinate Repo.preload in the wrong places, swallow exceptions with bare rescue, and skip Ecto.Multi for transactions. phxagents knows the patterns and refuses to ship code that breaks them.

Iron Laws enforced

26 non-negotiable rules: no Repo in loops, no bare rescues, no untyped Oban args, no missing preload. The judge agent blocks PRs that violate them.

iron-law-judge · /phx:review

Parallel code review, 4 specialists

A code review agent fans out to elixir-reviewer, security-analyzer, testing-reviewer and verification-runner — each reads the diff in fresh context, then findings are deduped and triaged. No more single-pass slop.

parallel-reviewer · /phx:review

Compound docs

Every fixed bug is captured as a searchable solution doc with YAML frontmatter. Next time you hit a similar Ecto constraint error, the answer is one query away.

/phx:compound

Workflow, not autocomplete

/phx:plan → /phx:work → /phx:review → /phx:compound. Each step has compile, format and test gates. Generated code that doesn't pass mix test doesn't get committed.

/phx:full · end-to-end

Context-aware

Reads mix.exs to discover credo, dialyzer, sobelow, ex_check. Honors your custom aliases. Knows when you're on Phoenix 1.7 vs 1.8 Scope. Doesn't dictate — adapts.

verification-runner agent

Supply-chain aware

Audits Hex deps for bidi chars, compile-time exec, maintainer changes, typosquats and CVEs. Records vetted versions in hex_vet.exs so you don't re-review the same package twice.

/phx:deps-audit · /phx:deps-vet

The workflow

Plan, work, review, compound.

One command per phase. Each phase produces an artifact the next phase reads. The cycle is the product.

  1. Plan

    Specialist agents draft a multi-step plan.md with files, schemas and Oban jobs to touch.

    /phx:plan

  2. Work

    Implements step-by-step. mix compile and mix test verified after each step.

    /phx:work

  3. Review

    4-track parallel review (Elixir, security, tests, verification) → findings → interactive triage.

    /phx:review

  4. Compound

    Captures what was learned as a searchable solution doc. Future you finds it in 2 seconds.

    /phx:compound

Supported clients

Shipping on Claude Code today. Three more in v3.0.

Claude Code is stable. Codex, OpenCode and Pi adapters land together in v3.0 with full Phase 1 + Phase 2 parity — single source, three generated targets. Track progress in PR #46.

Install in 30 seconds.

One plugin manifest in your .claude/ folder. Iron Laws and 50 skills auto-load.