# github-router

[![CI](https://github.com/animeshkundu/github-router/actions/workflows/ci.yml/badge.svg)](https://github.com/animeshkundu/github-router/actions/workflows/ci.yml)
[![npm](https://img.shields.io/npm/v/github-router)](https://www.npmjs.com/package/github-router)
[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)

Use your GitHub Copilot subscription to power **Claude Code**, **Codex CLI**, or any OpenAI/Anthropic-compatible tool.

github-router is a local reverse proxy that translates standard API formats to GitHub Copilot's backend. One command to start, copy-paste configs for your tools.

> [!WARNING]
> Unofficial. Not supported by GitHub. May break. Use responsibly.
> Review the [GitHub Copilot Terms](https://docs.github.com/site-policy/github-terms/github-terms-for-additional-products-and-features#github-copilot) and [Acceptable Use Policies](https://docs.github.com/site-policy/acceptable-use-policies/github-acceptable-use-policies#4-spam-and-inauthentic-activity-on-github).

## Quick Start

```sh
# 1. Authenticate (one-time)
npx github-router@latest auth

# 2. Start the proxy
npx github-router@latest start
```

The server runs at `http://localhost:8787`. Now pick your tool below.

---

## Use with Claude Code

**Option A: One-shot subcommand (recommended)**

```sh
npx github-router@latest claude
```

Boots the proxy on a random port and spawns Claude Code wired to it. Sets `ANTHROPIC_MODEL=claude-opus-4-7` (Anthropic's dashed slug — Claude Code's `/model` UI displays this as menu entry "Opus 4.7 (1M context)"). The proxy translates to Copilot's `claude-opus-4.7-1m-internal` on enterprise tokens or `claude-opus-4.7` on Pro+/Business/Max at request time. Major.minor fallback chain: `claude-opus-4-6` → `claude-opus-4-5`. Override with `-m`:

```sh
npx github-router@latest claude -m claude-opus-4-7
```

The launcher sanitizes parent-env auth keys and sets `CLAUDE_CONFIG_DIR=$HOME/.claude` so the spawned `claude` ignores any persisted Console OAuth credential without requiring `claude /logout`. Settings, MCP servers, hooks, and CLAUDE.md auto-discovery still load from `~/.claude` as normal.

**Option B: Interactive launch-command generator**

```sh
npx github-router@latest start --claude-code
```

Select your models, a launch command gets copied to your clipboard. Paste it in a new terminal.

**Option C: Copy-paste config**

Create `.claude/settings.json` in your project:

```json
{
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:8787",
    "ANTHROPIC_API_KEY": "dummy",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "gpt-4.1",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "gpt-4.1",
    "ANTHROPIC_SMALL_FAST_MODEL": "gpt-4.1-mini",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-4.1-mini",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "permissions": {
    "deny": ["WebSearch"]
  }
}
```

Then run `claude` as normal.

### Peer-MCP review subagents

`github-router claude` auto-wires four peer-model adversarial reviewers plus a coordinator into the spawned Claude Code session. No setup, no prior MCP config, no `.claude/agents/` files needed — they appear as Task `subagent_type` options the session can delegate to. Opt out with `--no-codex-mcp`.

Each persona is exposed both as a Claude Code subagent (callable via the `Task` tool) AND as an MCP tool at `mcp__gh-router-peers__<name>`. Personas are stateless: each invocation runs a fresh request against its model with a baked persona prompt — they have no access to your scrollback or project memory, so the lead must paste the artifact into the brief.

| Subagent | Model | Endpoint | Effort tiers (default) |
|---|---|---|---|
| `codex-critic` | gpt-5.5 | `/v1/responses` | low \| medium \| high \| xhigh (xhigh) |
| `codex-reviewer` | gpt-5.3-codex | `/v1/responses` | low \| medium \| high \| xhigh (xhigh) |
| `opus-critic` | claude-opus-4-7 | `/v1/messages` | low \| medium \| high \| xhigh (xhigh) |
| `gemini-critic` | gemini-3.1-pro-preview | `/v1/chat/completions` | low \| medium \| high (high) |
| `peer-review-coordinator` | (meta) | — | — |

`peer-review-coordinator` is a subagent (not an MCP tool) that fans out to the right combination of the four critics in parallel based on artifact type — plan, diff, single file, or long-context — and aggregates findings.

**Effort tiers** are exposed via the MCP tool's `effort` argument; subagents pass it through. All four tiers are accepted on every persona. `xhigh` routinely runs 60–90s; the proxy responds to `tools/call` requests with SSE-streamed responses (per MCP 2025-06-18 Streamable HTTP transport spec) so the connection stays open past the standard ~60s MCP per-tool-call ceiling and long calls complete transparently with no user setup.

`gemini-critic` only registers when `gemini-3.1-pro-preview` is present in your Copilot model catalog. If absent, the persona is silently dropped from both the MCP `tools/list` and the subagent set, and `peer-review-coordinator` skips it in routing decisions.

For codex-side write capability (a `codex-implementer` persona that can mutate files via Codex's tool-use sandbox), pass `--codex-cli`. Requires `codex` CLI 0.129+ on `PATH`; falls back to HTTP-only with a warning if codex is missing or older. Pass `--codex-mcp-only` to also pass `--strict-mcp-config` to Claude Code so only the proxy's MCP servers are loaded (hides any MCP servers in your existing `~/.claude/mcp.json`).

### Code search (`mcp__gh-router-peers__code_search`)

Alongside the peer reviewers, the same MCP surface exposes a `code_search` tool — fast structured code search over the workspace, ranked by **BM25F** (Robertson, Zaragoza, Taylor 2004) over four code-aware fields: matched line, surrounding context, file path tokens, and a symbol-definition heuristic. On top of that, the top hits get a tree-sitter pass that promotes true identifier-definition sites over incidental string matches; depth is controlled by the optional `structural` argument (`"full"` parses the top 50 hits, `"topN"` parses the top 10 for tighter latency on big repos). A single `notice` field surfaces in the response on the rare occasions an actionable degradation fires — the structural pass overran its 200ms wall-clock budget, or the response hit the 256KB size cap and was truncated; the message text tells the model what to retry. Defaults to a "ranked" mode with shoulder pruning so models get the few right answers, not a flood of substring matches. `literal` and `regex` modes are also available for exact searches; single-identifier queries in `ranked`/`literal` mode auto-expand across camelCase / snake_case / kebab-case skeletons so `getUserName` also matches `get_user_name`.

`workspace` is any absolute path the proxy process can read — typically the project root or a sub-tree you're working in. The model picks it. There's no allow-set or secret-shape file denylist: the threat model is symmetric since Claude Code already has Read / Bash / Edit tools that reach the same paths, so gating one tool would have been inconsistency rather than defense. Paths in results are returned relative to the workspace, never absolute.

Ripgrep is provided via the `@vscode/ripgrep` npm dependency (per-platform binary via `optionalDependencies` — no postinstall script needed). The proxy prefers system `rg` on `PATH` when available and falls back to the bundled binary otherwise. To opt into raw query/path logging for debugging, set `GH_ROUTER_DEBUG_CODE_SEARCH=1` — by default the proxy logs only counts and timings.

---

## Use with Codex CLI

The fastest path is the `codex` subcommand — it boots the proxy on a random port and spawns Codex CLI wired to it:

```sh
npx github-router@latest codex
```

Defaults to `gpt-5.5`; falls back to `gpt-5.4` → `gpt-5.3-codex` → `gpt-5.2-codex` if your Copilot tier doesn't expose 5.5 yet. Override with `-m`:

```sh
npx github-router@latest codex -m gpt-5.3-codex
```

Or run the proxy and Codex CLI separately:

```sh
npx github-router@latest start --codex   # interactive launch-command generator
# — or set env vars yourself —
export OPENAI_BASE_URL="http://localhost:8787/v1"
export OPENAI_API_KEY="dummy"
codex --full-auto -m gpt-5.5
```

---

## Use with any OpenAI-compatible tool

Point any tool at `http://localhost:8787/v1`:

```sh
curl http://localhost:8787/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-4.1", "messages": [{"role": "user", "content": "Hello"}]}'
```

---

## API Endpoints

| Endpoint | Method | Format |
|---|---|---|
| `/v1/chat/completions` | POST | OpenAI Chat Completions |
| `/v1/responses` | POST | OpenAI Responses (Codex models) |
| `/v1/messages` | POST | Anthropic Messages |
| `/v1/messages/count_tokens` | POST | Anthropic token counting |
| `/v1/models` | GET | OpenAI model list |
| `/v1/embeddings` | POST | OpenAI embeddings |
| `/v1/search` | POST | Web search |
| `/usage` | GET | Copilot usage & quotas |

OpenAI-compatible endpoints are also available without the `/v1` prefix (for example, `/chat/completions`).
Anthropic endpoints are only available under `/v1/messages`.

<details>
<summary>Model / endpoint compatibility</summary>

| Model | /chat/completions | /responses | /v1/messages |
|---|---|---|---|
| gpt-4.1, gpt-4o | Yes | Yes | No |
| gpt-5.5, gpt-5.4 | No | Yes | No |
| gpt-5.3-codex, gpt-5.2-codex | No | Yes | No |
| claude-opus-4.7-1m-internal (enterprise), claude-opus-4.7 | Yes | No | Yes |
| claude-opus-4.6-1m, claude-opus-4.6, claude-sonnet-4.6 | Yes | No | Yes |
| o3, o4-mini | Yes | Yes | No |

</details>

---

## Docker

Pre-built images on GitHub Container Registry:

```sh
docker pull ghcr.io/animeshkundu/github-router:latest
docker run -p 8787:8787 -e GH_TOKEN=your_token ghcr.io/animeshkundu/github-router
```

Or build locally:

```sh
docker build -t github-router .
docker run -p 8787:8787 -e GH_TOKEN=your_token github-router
```

<details>
<summary>Docker Compose</summary>

```yaml
services:
  github-router:
    build: .
    ports:
      - "8787:8787"
    environment:
      - GH_TOKEN=your_github_token_here
    restart: unless-stopped
```

</details>

<details>
<summary>Persistent token storage</summary>

```sh
mkdir -p ./github-router-data
docker run -p 8787:8787 -v $(pwd)/github-router-data:/root/.local/share/github-router github-router
```

</details>

---

## CLI Reference

```
github-router start [options]    Start the proxy server
github-router claude [options]   Start proxy + spawn Claude Code wired to it
github-router codex [options]    Start proxy + spawn Codex CLI wired to it
github-router auth               Authenticate with GitHub
github-router check-usage        Show Copilot usage/quotas
github-router debug              Print diagnostic info
```

The `claude` and `codex` subcommands accept all the shared flags below plus `-m`/`--model` to override the default model. Default models live in `src/lib/port.ts`:

- `claude` → `claude-opus-4-7` (Anthropic dashed slug for UI compatibility; the proxy translates to Copilot's `claude-opus-4.7-1m-internal` on enterprise or `claude-opus-4.7` elsewhere). Major.minor fallback chain: `claude-opus-4-6` → `claude-opus-4-5`.
- `codex` → `gpt-5.5` → `gpt-5.4` → `gpt-5.3-codex` → `gpt-5.2-codex`

Fallback chains fire only on the implicit-default path; explicit `-m`/`--model` is always respected as-is.

| Flag | Description | Default |
|---|---|---|
| `--port, -p` | Port | 8787 |
| `--verbose, -v` | Debug logging | false |
| `--account-type, -a` | `individual` / `business` / `enterprise` | individual |
| `--rate-limit, -r` | Min seconds between requests | - |
| `--wait, -w` | Queue requests instead of rejecting on rate limit | false |
| `--manual` | Approve each request manually | false |
| `--github-token, -g` | Pass token directly (skip auth flow) | - |
| `--claude-code, -c` | Generate Claude Code launch command | false |
| `--codex` | Generate Codex CLI launch command | false |
| `--show-token` | Print tokens to console | false |
| `--proxy-env` | Use HTTP_PROXY/HTTPS_PROXY env vars | false |

Additional flags accepted only by the `claude` subcommand:

| Flag | Description | Default |
|---|---|---|
| `--model, -m` | Override the default Claude model | claude-opus-4-7 |
| `--codex-mcp` / `--no-codex-mcp` | Wire peer-MCP review subagents (codex-critic / opus-critic / gemini-critic / codex-reviewer / peer-review-coordinator) into the spawned session | true |
| `--codex-cli` | Add a `codex mcp-server` stdio backend so `codex-implementer` can mutate files. Requires codex CLI 0.129+; falls back to HTTP-only if absent | false |
| `--codex-mcp-only` | Pass `--strict-mcp-config` to Claude Code so only the proxy's MCP servers load (hides any user MCP servers in `~/.claude/mcp.json`) | false |
| `--stealth` | Opt back into VS Code-only beta-header filtering. Loses leverage features (task budgets, token-efficient tools, prompt caching, etc.) but minimizes the wire-fingerprint diff from VS Code Copilot Chat | false |
| `--auto-update` / `--no-auto-update` | Check for and install latest Claude Code on launch (throttled to once per hour). Falls back gracefully if npm/network unavailable | true |
| `--update-check` / `--no-update-check` | Check the npm registry for a newer Claude Code version on launch and warn if stale (~500ms cost). `--no-update-check` implies no auto-install | true |

---

## Development

```sh
bun install          # Install deps
bun run dev          # Dev server with hot reload
bun test             # Run tests
bun run lint:all     # Lint
bun run typecheck    # Type check
bun run build        # Build for distribution
```

## License

[MIT](LICENSE)