DOCS

Roadmap

Next steps for GateFox and where contributors can help: the Claude apps, DeepSeek, and more harnesses.

About 4 min read · 6 sections

On this page

GateFox already serves Claude, ChatGPT, Copilot, OpenRouter, xAI, OpenCode and local models to Codex, Claude Code and Claude Desktop. The next steps make the Claude apps feel native, add direct DeepSeek support, and open the gateway to any harness that accepts an API endpoint, and bring it to Linux and Windows.

Each item links to its GitHub issue; all of them carry the roadmap label. Comment there before you start, so two people do not build the same thing. Items marked good first issue need no deep knowledge of the gateway. Read CONTRIBUTING.md first.

            today                                  next
 Codex (Mac + phone)┐                  ┌──────────── Cline · Aider · Zed · Continue
 Claude Code ───────┼─► GateFox ◄──────┤             (Chat Completions ingress)
 Claude Desktop ────┘   │  macOS       └──────────── Linux · Windows
                        ▼
   Claude · ChatGPT · Copilot · OpenRouter · xAI · local    + DeepSeek direct

1. A better Claude app experience

Work Why Start in
Report each route's real context window to Claude Desktop (#73) Routes show Claude's generic 200K or 1M estimate, so compaction can start too late or too early Claude Desktop, src/claude-desktop-*.ts
Set up Claude Desktop from the GateFox app (#74) Today it needs commands and a manual quit and apply Claude Desktop, app management
Native reasoning and Fast controls without patching Claude (#75) The native picker trial works only on a patched copy of one Claude version scripts/claude-native-trial/
Long-session checks for Claude Code (#76) Short checks do not cover long compaction or every MCP schema Messages, scripts/e2e-*.mjs
Clearer errors for unsupported Claude features (#77) A route that cannot honour a feature should say so in the client, not fail later Known request rejections

2. DeepSeek

DeepSeek works today through OpenRouter. A direct connection was built and later removed on 6 November 2026; it can come back as a plain API preset.

Work Why Start in
Direct DeepSeek API preset (#78) Lower cost and latency than going through OpenRouter, with DeepSeek's own prefix caching API providers, src/providers/
Reasoning replay and cache usage for DeepSeek (#79) Keep reasoning across tool calls and report cache hits in the usage ledger Cache safety
DeepSeek in the app's Add connection wizard (good first issue) (#80) Today it needs a JSON edit App management

3. Any harness that takes an API endpoint

GateFox accepts the OpenAI Responses and Anthropic Messages APIs. Many coding harnesses only speak Chat Completions.

Work Why Start in
/v1/chat/completions ingress, with streaming and tool calls (#81) Unlocks Cline, Aider, Zed, Continue, Goose and most other harnesses src/server.ts, src/protocol/, architecture
/v1/models listing for that ingress (good first issue) (#82) Harnesses fill their model picker from it src/catalog.ts
Setup recipes per harness (good first issue) (#83) One short, tested page each: OpenCode, Cline, Aider, Zed, Continue, Goose new docs/clients/<name>.md
A harness checklist What to verify before calling a harness supported Checklist for a new harness

The gateway translates and transports; the harness keeps its tools, permissions and model choices. New ingress must keep that rule (architecture).

4. More platforms

GateFox runs on macOS (Apple silicon). Its core is Node, so most of it can move.

Platform Today Help wanted
macOS Full support: app, background service, updater, Codex and Claude desktop apps —
Linux The command-line gateway runs; no published package Linux archive and install guide (#84)
Windows Not supported yet Paths, private credential files, a Windows service, and desktop app registration (#86)
Phone Codex in the ChatGPT app lists and continues gateway chats through your Mac's native relay Claude's mobile Remote Control does not accept custom gateways yet

5. Easier to start

Work Why
A short demo video or GIF in the README (good first issue, #85) Shows the model picker with added models in a few seconds

Proposing something else

Open an issue that says who it helps and what you would change. Larger changes are easier to review when we agree on the approach first.