Codex-native specialist routing

Bring a specialist model into Codex.
Keep Codex in control.

For developers who want a faster second opinion on code, architecture, tests, or a bounded implementation proposal—without switching tools or giving another model repository access.

Install the plugin See how it works
  • Review without tab switching
  • Bounded evidence and spend
  • No automatic writes
Kimi Qwen Claude

Free · bring your own API key

Add a specialist model to Codex in minutes.

Clone the plugin, add its local marketplace to Codex, and keep OpenRouter or direct-provider credentials in the operating system keyring.

git clone https://github.com/web5labs/empire-llm-codex cd empire-llm-codex codex plugin marketplace add "$PWD" codex plugin add empire-llm-codex@empire-local

Get Empire on GitHub Explore Empire LLM
0 leadCodex owns repository decisions
0 specialistBest-fit model for the task
0 budgetSpend authorized before the call
0 silent writesExternal output stays advisory
Kimi Qwen Gemini Grok Claude OpenRouter

Focused by design

Use another model where it saves time—not everywhere.

Empire routes one bounded job to a current specialist, normalizes the response, and returns it to Codex with cost and provenance attached.

01 / Review

Catch issues before another edit cycle

Ask a strong external model to challenge the selected diff. Codex checks each finding against your repository before changing anything.

  • Selected files only
  • Normalized tables
  • Model and cost provenance

02 / Handoff

Turn a spec into a reviewable starting point

Generate a checklist or one proposed source file outside the repository. Codex previews, validates, and decides what is worth implementing.

  • Hash and size validation
  • Source-only preview
  • Approval never applies code

03 / Control

Know what was sent, served, and spent

See the selected evidence, model identity, routing rationale, latency, and estimated cost with every contribution.

  • OS system keyring
  • Local budget ledger
  • Clear missing-evidence states

Built for focused builders

More perspective. Less model switching.

Empire is for developers who want another model’s strengths without adding another coding environment, autonomous agent, or review queue.

01

Solo developers

Pressure-test decisions and catch regressions without manually copying context between tools.

02

Small engineering teams

Add a second perspective while keeping repository access and implementation authority with Codex.

03

Agentic builders

Match reviews, specs, and bounded proposals to models that are strong at that specific job.

Five explicit steps

A faster path from uncertainty to verified action.

  1. 01

    Select

    Codex bounds the task, diff, or files.

  2. 02

    Route

    Empire scores eligible models for the job.

  3. 03

    Contribute

    One model returns an advisory result.

  4. 04

    Verify

    Codex checks claims against local evidence.

  5. 05

    Ship

    You approve what Codex implements.

Interactive model lab

See how a routing profile changes the contribution.

This illustrative interface explains routing dimensions. It does not claim a live ranking.

Worker route

Kimi K3

moonshotai/kimi-k3

9.4/10
Task fitExcellent
EfficiencyFocused
EvidenceBounded
Example synthesis< $0.10 ceiling

Kimi contributes a front-end proposal. Codex verifies accessibility, repository fit, and implementation details before writing any project file.

Illustrative · live catalog evaluated at request time

Local control

Your keys stay in the system keyring. Your code stays bounded.

Empire reads credentials from macOS Keychain, Windows Credential Manager, or Linux Secret Service. Keys never belong in chat, repositories, logs, fixtures, or routed responses.

Read the security model
  • Environment or system keyringNever plaintext settings
  • Eight files maximum80 KB default evidence ceiling
  • OpenAI excluded by defaultCodex is already the OpenAI lead
  • Advisory by constructionNo external repository permissions