Frontier

Frontier

Frontier is an orchestration runtime for frontier harnesses and models that need to delegate bounded work to other harnesses, models, and providers. It starts with Claude Code and Codex plugin surfaces plus a deterministic companion CLI that routes work through harness adapters such as Pi and Codex.

Status: early public preview. The runtime is usable, but provider support is still evolving.

The working pattern is simple: keep planning, tradeoffs, synthesis, and final review with the frontier orchestrator; delegate bounded research, implementation, testing, and log reduction to configured worker harnesses. Provider mechanics are configuration, never baked into prompts.

Why Frontier

Frontier exists because model capability and harness ergonomics do not always ship together. You may prefer Claude Code, Codex, or another frontier harness as your working environment, while wanting to delegate bounded work to models that harness does not offer: local oMLX models, Ollama models, OpenAI-compatible gateways, or cheaper cloud models. Frontier separates orchestration from model access so the harness you like can coordinate the models you choose.

It also exists because frontier models are becoming simultaneously more capable, more expensive, and more scarce. Agentic coding does not spend tokens only on final reasoning; it burns context on search, retries, logs, dependency inspection, implementation attempts, and review loops.

That cost pattern is measurable. A 2026 study of agentic coding workloads found that agentic tasks can consume 1000x more tokens than ordinary code chat or code reasoning, with same-task runs varying by up to 30x and higher token usage not reliably improving accuracy: How Do AI Agents Spend Your Money?

The market pressure is visible in current frontier releases. Claude Fable 5 was reported as Anthropic's most capable broadly available model, but with pricing at $10 per million input tokens and $50 per million output tokens, plus safeguards that can route sensitive requests to another model: The Verge

Teams are already responding to this economics problem. Reporting on "token-maxxing" describes founders and companies adding spend visibility, approval flows, and token caps after surprise AI-agent costs: Business Insider

Frontier's answer is orchestration. Keep your preferred frontier harness focused on planning, judgment, and final review; route bounded scans, summaries, implementation attempts, and log reduction to cheaper, local, or specialized worker harnesses and models.

Get Started

Install Frontier from its Claude Code marketplace:

/plugin marketplace add WellDunDun/frontier
/plugin install frontier@frontier-marketplace

Or paste this into an agent with terminal access:

Install Frontier for Claude Code. Run these commands:

claude plugin marketplace add WellDunDun/frontier --scope user --sparse .claude-plugin plugins
claude plugin install frontier@frontier-marketplace --scope user
claude plugin list

Then stop and tell me whether Frontier is installed. Do not run provider setup
yet; I will restart Claude Code if needed and run /frontier:setup myself.

Then configure the worker backend:

1. Choose and start or configure a backend:

   • oMLX: start the server with omlx start or omlx serve <model>.
   • Ollama: start Ollama and make a model available.
   • OpenAI-compatible local or cloud backend: add frontier.config.json with flavor: "openai", baseUrl, and an apiKey.env reference when auth is required.

   omlx launch <tool> opens an interactive configure-and-launch TUI; run it yourself if useful. Frontier does not invoke it for you.

2. Check readiness:

   /frontier:setup
   

3. Provision the Pi provider and Codex profile if they are missing:

   /frontier:setup --apply
   

   The apply step provisions both harnesses additively: it ensures the selected Pi provider exists in ~/.pi/agent/models.json and the selected Codex provider/profile exist in ~/.codex/config.toml, taking a timestamped backup before any write and preserving unrelated entries. Models are read from the configured backend, so the backend must be reachable when you apply.

4. Smoke test delegation:

   /frontier:delegate reply with OK
   

The checkout-based --plugin-dir flow is only needed for local development.

Examples

Delegate a research or review task through the default worker harness: