copilot-bridge

Use GitHub Copilot as a local OpenAI/Anthropic-compatible API, so Codex CLI, Claude Code and Continue can talk to Copilot with minimal configuration.

[!WARNING]
This is a reverse-engineered bridge for the GitHub Copilot API. It is not
supported by GitHub, and may break unexpectedly. Use at your own risk.

[!WARNING]
GitHub Security Notice:
Excessive automated or scripted use of Copilot (including rapid or bulk
requests, such as via automated tools) may trigger GitHub's abuse-detection
systems. You may receive a warning from GitHub Security, and further
anomalous activity could result in temporary suspension of your Copilot
access.

GitHub prohibits use of their servers for excessive automated bulk activity
or any activity that places undue burden on their infrastructure.

Please review:

• GitHub Acceptable Use Policies
• GitHub Copilot Terms

Use this bridge responsibly to avoid account restrictions. The
--rate-limit <seconds> flag is provided to help throttle upstream traffic.

Demo

Codex CLI

Claude Code

Install & run

# one-time GitHub device login
npx betahi-copilot-bridge@latest auth

# start the bridge on 127.0.0.1:4142
npx betahi-copilot-bridge@latest start

start flags: --host, --port, --show-token, --debug,
--no-codex-setup, --select-model, --no-prompt,
--rate-limit <seconds>, --wait.

After startup the banner prints a Usage Viewer link of the form
https://betahi.github.io/copilot-bridge?endpoint=http://127.0.0.1:4142/usage,
which renders the Copilot quota snapshot (chat / completions / premium
interactions) read from GET /usage.

The bridge exposes both adapter-style endpoints (/v1/responses,
/v1/messages) and the raw OpenAI-compatible surface
(/v1/chat/completions, /v1/embeddings, /v1/models) so tools like
LiteLLM, Continue, Cline and Aider work out of the box. CORS is enabled
globally for browser-based clients.

--rate-limit N enforces a minimum of N seconds between upstream
requests (anti–abuse-detection throttle). Add --wait to block instead
of returning HTTP 429 when the window has not elapsed.

--debug enables extra upstream error diagnostics in console logs.

• Includes: token limits, stream mode, tool count, invalid tool names, and suspicious tool-schema paths.
• Does not include: request messages, prompt text, bearer tokens, tool descriptions, or the full request body.

Review or redact debug logs before sharing them publicly.

Configure Codex CLI

start writes a managed block into ~/.codex/config.toml. You don't edit
this block; the bridge regenerates it on every start.

# >>> copilot-bridge managed block — auto-generated, do not edit between markers >>>
model_provider = "bridge"

[model_providers.bridge]
name = "Copilot Bridge"
base_url = "http://127.0.0.1:4142/v1"
wire_api = "responses"
prefer_websockets = false
requires_openai_auth = false
# <<< copilot-bridge managed block — edits outside this block are preserved <<<

To pin the default model for codex (without passing -m every time),
edit the top of the same file — outside the markers. The bridge preserves
these keys across rewrites:

model = "gpt-5.3-codex"
model_reasoning_effort = "high"

If model_reasoning_effort is omitted, the bridge leaves it unset and does not
send a reasoning effort for Codex requests (a startup info log also reminds you).

model_reasoning_effort only affects Codex requests; it does not affect Claude.

That's it — codex exec '...' will now route through the bridge to Copilot.

Use --no-codex-setup to skip the managed-block writer entirely (e.g. if you
manage ~/.codex/config.toml yourself).

Codex warning: "Model metadata ... not found"

This is a Codex client-side metadata warning, not a bridge routing failure.
Requests can still complete through the bridge.

For claude-opus-4.6-1m, upstream still enforces a 1,000,000-token prompt
limit (about 900k succeeds; around 1,000,046 is rejected as too long).

Configure Claude Code

Bridge endpoints: POST /v1/messages, POST /v1/messages/count_tokens.

copilot-bridge start automatically writes ANTHROPIC_BASE_URL (and a dummy
ANTHROPIC_AUTH_TOKEN if absent) into ~/.claude/settings.json, so
claude works the moment the bridge is listening — even on a different host
or port. Pass --no-claude-setup to skip this. All other keys in your
settings file (model overrides, plugins, marketplaces) are preserved.

Minimal recommended config is:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:4142",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "claude-opus-4.7",
    "MODEL_REASONING_EFFORT": "medium"
  }
}

Optional slot-specific:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:4142",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "claude-opus-4.7",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4.6",
    "ANTHROPIC_SMALL_FAST_MODEL": "claude-haiku-4.5",
    "MODEL_REASONING_EFFORT": "medium"
  }
}

MODEL_REASONING_EFFORT (case-insensitive key lookup) is also read from the
project-local .claude/settings.json and .claude/settings.local.json and
applied to Claude requests only when the model supports reasoning. If it is not
configured, Claude requests do not infer or attach a reasoning effort.

Environment overrides

Supported models

The bridge resolves aliases and clamps reasoning effort to what each model
accepts upstream.

GPT-5 family — native Responses passthrough

Claude family — translated to chat completions

Gemini family — translated to chat completions

Legacy

gpt-4.1, gpt-4o — chat-only upstream, no reasoning parameter.

Reasoning effort

For OpenAI-compatible clients, unsupported reasoning values are clamped to the
model capability table instead of being forwarded upstream. If a request omits
reasoning effort, the bridge leaves it omitted rather than inferring a default.
Claude-side reasoning can be set globally via MODEL_REASONING_EFFORT (env, or
env in ~/.claude/settings.json); invalid Claude-side values are ignored,
and per-request reasoning_effort takes precedence.

Development

Requires Bun ≥ 1.2.

bun install
bun run dev          # watch mode against src/main.ts
bun test             # run all tests (bun test runner)
bun run typecheck    # tsc --noEmit
bun run build        # produce dist/main.js with tsdown

# run directly from source (no build, no npx); --port specifies the port
bun run ./src/main.ts start --host 127.0.0.1 --port 4141 --no-prompt

Adding another CLI: drop a new translator under src/bridges/<client>/,
reuse src/services/copilot/ for upstream calls, register routes in
src/server.ts, and add tests under tests/.

License

MIT — see LICENSE.