Frontier PROXY

Documentation

Overview

Frontier Proxy is a desktop orchestrator that routes coding tasks to the CLI agents already installed and signed in on your machine.

What it is

Frontier Proxy is an Electron desktop app that sits in front of the coding agents you already use — Codex CLI, Claude Code, GitHub Copilot CLI, and Ollama-backed models. You describe a task once; Frontier decides which agent should take it, runs that agent locally, streams the result back, and keeps the conversation, file changes, and usage in one place.

How a task flows

  1. You describe the task

    Prompt, working directory, and routing mode. Optionally pin a specific agent and model, attach images, or reference files and folders with @.

  2. Frontier classifies and queues it

    The task is typed as coding, debugging, review, planning, documentation, or general work, then queued respecting global and per-provider concurrency limits.

  3. A provider is ranked and chosen

    Eligible agents are scored on your override, the routing policy, task affinity, your priority order, and current usage and load. Anything rate-limited, over budget, or cooling down is skipped.

  4. The CLI runs on your machine

    The agent is spawned directly — never through a shell — with the prompt on stdin. Output streams back with tool calls, thinking, and the files it touched.

  5. You follow up

    Reply in the same thread, retry, cancel, or hand the conversation to a different agent. On quota errors Frontier fails over on its own.

Frontier Proxy architecture The desktop UI sends tasks through a secure bridge into the scheduler and router, which launch local Codex, Claude Code, Codex with Ollama, or Ollama processes. DESKTOP UI tasks · output providers · limits SECURE IPC Orchestration core CLASSIFY + SCORE QUEUE + CONCURRENCY USAGE · COOLDOWN · FAILOVER STDIN / JSONL Codex CLI Claude Code Codex + Ollama Ollama Custom local CLI NO MODEL API KEYS · NO SHELL INTERPOLATION

What it deliberately is not

  • Not an API client. Frontier proxies CLI processes, not the private internals or UI automation of the Codex or Claude desktop apps. That is the stable local integration surface, and it still reuses the account sessions those CLIs already hold.
  • Not a billing dashboard. Subscription tools expose no reliable universal "tokens remaining" interface. Frontier shows what each CLI actually reports, tracks conservative local estimates for the rest, and never invents a percentage.
  • Not a sandbox. Every agent can modify its selected working directory. Review the permission mode, work in a repository, and use the Review screen before merging anything.

Where to next

Install

Download a build or compile from source.

Provider setup

Register and sign in to Codex, Claude Code, Copilot, and Ollama.

Routing & failover

How a provider is chosen and when work reroutes.

Context & Tools

One MCP and tool profile translated into every CLI's flags.

Orchestration & bench

Split work across agents, or race them head to head.

Task workspace

Conversations, file changes, context meter, and usage.

Security & data

What runs where, and what never leaves your machine.

Troubleshooting

Fixes for the failures people actually hit.

Prefer to read the source? Start with security & data, then the repository itself.