Local Codex for Enterprise

The governed, local-first AI engineering platform for small teams.

Give your team governed access to local coding agents — without shipping your code, prompts, or model traffic to a third party. Local Codex for Enterprise is a self-hosted control plane with delegated multi-agent execution, immutable audit evidence, and a library-backed knowledge system. It runs on a laptop for evaluation and deploys to a server for the team.

Not affiliated with, endorsed by, or supported by OpenAI. Derived from the open-source OpenAI Codex project; Apache-2.0 and upstream attribution are preserved.

Why teams choose it

• Governed by default. Seeded roles, workspace allowlists, fail-closed provider/model/worker authority, append-only audit events and execution receipts. No hidden privilege escalation, no unrecorded shared mutation.
• Local-first inference. Choose Ollama (local), Docker Model Runner (local), or OpenAI-compatible remote. On macOS the recommended default is Ollama; on other platforms Ollama and Docker are equal options. Nothing leaves your machine unless you configure a remote provider.
• Built for delegated work. Validated at 4 users, 4 delegated agents per user, and 20 concurrent agent lifecycles (locally demonstrated plus a deterministic simulation) — with bounded queues, admission control, cancellation, restart recovery, and truthful queued → running → completed / cancelled / failed states. Admission limits are configuration, not a fixed product ceiling.
• Knowledge that stays scoped. Upload files and LCFE-created artifacts to the Library; retrieve them through the governed library.search MCP tool with project-scoped pgvector results that remain attributable and auditable.
• Honest about its own beta. The scheduler defaults to deterministic smoke mode; real model-backed chat is verified locally via Ollama. No simulated state is presented as observed fact.

Screenshots

Captured against the rebuilt local demo (Ollama-backed chat, fixed workbench shell):

Try it in three steps

1. Install Ollama (macOS default provider):

brew install ollama
ollama serve
ollama pull qwen2.5:1.5b-ctx32

2. Start the control plane:

LOCAL_CODEX_ENTERPRISE_WORKSPACES="$PWD" \
LOCAL_CODEX_ENTERPRISE_HANDOFF_TOKEN_SECRET="change-me-for-real-use" \
  docker compose -f compose.enterprise.yaml up --build

3. Open http://127.0.0.1:8787, create your owner via /setup, choose Ollama (local) in Admin → Settings, and send your first chat message.

Providers

No provider is auto-seeded: after /setup, an owner/admin saves the provider at /admin/settings. New worker starts fail closed until an approved provider/model exists. Ollama setup is install-only — LCFE detects it from the settings page and shows brew install ollama / ollama serve guidance when it is unreachable; no model auto-pull, no secrets stored (API keys are referenced by environment-variable name only).

Status

Current beta: 0.0.1-beta.10.1. The candidate is verified end-to-end — governed library search, upload/download round-trip, scheduler reliability, restart and recovery, queue saturation and admission, cancellation under load, mutation isolation, multi-agent capacity, real model-backed chat via Ollama, the full test suite, and settings SQL export/import. See the engineering guide for the honest limits and deployment details.

For engineers

• Engineering guide — deployments, environment reference, architecture, providers, context packs, scheduled sessions, security, and current limits.
• Domain contract · Context Pack contract · Roadmap
• Security · Threat model · Contributing

License

Apache-2.0, derived from OpenAI Codex with upstream attribution in NOTICE.