claude-code-proxy lets you use Claude Code with your ChatGPT Plus/Pro subscription or your Kimi Code (kimi.com) account.

Quick start · Providers · How it works · Configuration · Limitations

Why?

I feel Claude Code is still the best harness around, despite occasional frustrations caused by updates. However, Anthropic keeps tightening the usage limits, while OpenAI is still much more generous.

If you want to use OpenAI plans, your best options seem to be OpenCode and Codex. I tried OpenCode, but the UX has many rough edges, especially around skills feeling like a second-class feature. Fortunately it's open source and I ended up forking it and applying some patches, but would much rather not do it.

Quick start

1. Install

Homebrew (macOS and Linux):

brew install raine/claude-code-proxy/claude-code-proxy

Install script (macOS and Linux):

curl -fsSL https://raw.githubusercontent.com/raine/claude-code-proxy/main/scripts/install.sh | bash

Manual: download a prebuilt binary for your platform from the releases page. Windows artifacts are published as claude-code-proxy-windows-amd64.zip and claude-code-proxy-windows-arm64.zip; extract the .exe somewhere on your PATH.

2. Pick a provider and authenticate

The proxy supports two upstream providers. Pick one and run its login flow; the proxy will refuse to start traffic until a token is stored.

Codex (ChatGPT Plus/Pro):

claude-code-proxy codex auth login     # browser OAuth (PKCE)
# or, on a headless machine:
claude-code-proxy codex auth device    # device-code flow

Sign in with your ChatGPT Plus/Pro account, not an OpenAI API account.

Kimi (kimi.com Kimi Code):

claude-code-proxy kimi auth login      # device-code flow (prints URL + code)

Sign in with your kimi.com account. The verification URL is displayed; open it in any browser, confirm the code, and the CLI polls until done.

On macOS credentials go to Keychain. On Windows they are written under %APPDATA%\claude-code-proxy\<provider>\auth.json; on Linux they are written under ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/<provider>/auth.json (mode 0600 where supported).

Verify:

claude-code-proxy codex auth status
claude-code-proxy kimi auth status

3. Start the proxy

claude-code-proxy serve                # listens on 127.0.0.1:18765
PORT=11435 claude-code-proxy serve     # change the listen port

Binds to 127.0.0.1 only. One serve process handles all providers — the upstream for each request is chosen from ANTHROPIC_MODEL.

4. Point Claude Code at it

ANTHROPIC_MODEL selects the provider:

• gpt-5.5, gpt-5.4, gpt-5.3-codex, gpt-5.3-codex-spark, gpt-5.4-mini, gpt-5.2 → codex
• kimi-for-coding, kimi-k2.6, k2.6 → kimi

An unknown model returns a 400 listing the supported ids. There is no implicit default provider.

Claude Code also issues background requests (session title generation, token counts) against its built-in "small/fast" haiku model id. Those requests would 400 because no provider claims it, so set ANTHROPIC_SMALL_FAST_MODEL to a concrete id too (the same value as ANTHROPIC_MODEL is usually fine):

# Codex
ANTHROPIC_BASE_URL=http://localhost:18765 \
ANTHROPIC_AUTH_TOKEN=unused \
ANTHROPIC_MODEL=gpt-5.5[1m] \
ANTHROPIC_SMALL_FAST_MODEL=gpt-5.4-mini[1m] \
CLAUDE_CODE_AUTO_COMPACT_WINDOW=272000 \
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 \
  claude

# Kimi
ANTHROPIC_BASE_URL=http://localhost:18765 \
ANTHROPIC_AUTH_TOKEN=unused \
ANTHROPIC_MODEL=kimi-for-coding[1m] \
ANTHROPIC_SMALL_FAST_MODEL=kimi-for-coding[1m] \
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 \
CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 \
  claude

CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1 is recommended because the proxy always talks to upstream providers with streaming requests, even when it accumulates a non-streaming Anthropic response for Claude Code. Disabling Claude Code's streaming-to-non-streaming fallback avoids retrying a partially completed stream in a way that can duplicate tool calls.

Or set it persistently in ~/.claude/settings.json:

{
  "env": {
    "ANTHROPIC_BASE_URL": "http://127.0.0.1:18765",
    "ANTHROPIC_AUTH_TOKEN": "unused",
    "ANTHROPIC_MODEL": "gpt-5.5[1m]",
    "ANTHROPIC_SMALL_FAST_MODEL": "gpt-5.4-mini[1m]",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": 272000,
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": 1,
    "CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK": 1
  }
}

5. Context window size

Claude Code decides auto-compaction based on the model's context window. For unknown models, Claude Code uses its own fallback context size. The [1m] suffix is a local Claude Code hint that raises that compaction threshold. It is useful only when the upstream model can actually handle a window that large.

Use the [1m] suffix for Codex and Kimi models so Claude Code uses a larger local compaction threshold, such as gpt-5.5[1m], gpt-5.4-mini[1m], or kimi-for-coding[1m]. The proxy strips a trailing [1m] before sending the request upstream. The suffix affects Claude Code's local compaction decision and does not increase the upstream model's context window.

Official Codex metadata reports gpt-5.5 with a 272K-token window, not a 1M window. Set CLAUDE_CODE_AUTO_COMPACT_WINDOW=272000 with gpt-5.5[1m] so Claude Code does not compact too early, while still compacting before the real upstream limit.

If you'd rather disable auto-compact completely, set DISABLE_AUTO_COMPACT=1 in your env or ~/.claude/settings.json. Manual /compact still works, but you risk hitting real upstream limits before Claude Code can compact for you.

Toggling between proxy and direct Anthropic

If you still have an Anthropic subscription you want to fall back to, you can put a small wrapper in front of claude that only injects the proxy env vars when a flag file exists, plus a toggle script to flip the flag. Leave ~/.claude/settings.json free of proxy env vars so direct-to-Anthropic remains the default.

~/.local/bin/claude (ahead of the real claude on PATH):

#!/bin/bash
# Wrapper that optionally routes to claude-code-proxy.
# Active when ~/.claude/claude-code-proxy-enabled exists.

if [ -f "$HOME/.claude/claude-code-proxy-enabled" ]; then
    export ANTHROPIC_BASE_URL="http://localhost:18765"
    export ANTHROPIC_AUTH_TOKEN="unused"
    export ANTHROPIC_MODEL="gpt-5.5[1m]"
    export ANTHROPIC_SMALL_FAST_MODEL="gpt-5.4-mini[1m]"
    export CLAUDE_CODE_AUTO_COMPACT_WINDOW="272000"
    export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC="1"
    export CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK="1"
fi

exec "$HOME/.local/bin/claude" "$@"

Adjust the exec path if the real claude binary lives elsewhere on your system (e.g. $(bun pm bin -g)/claude, $HOME/.claude/local/claude).

claude-proxy-toggle (anywhere on your PATH):

#!/bin/bash
# Toggle claude-code-proxy routing for the claude wrapper.
set -euo pipefail

flag="$HOME/.claude/claude-code-proxy-enabled"

if [ -f "$flag" ]; then
    rm "$flag"
    echo "proxy: off"
else
    mkdir -p "$(dirname "$flag")"
    touch "$flag"
    echo "proxy: on"
fi

Run claude-proxy-toggle to flip between routing through the proxy (Codex / Kimi) and talking to Anthropic directly. New or continued claude sessions pick up the change immediately; existing sessions keep whatever they started with.

Providers

Codex (ChatGPT)

Upstream: https://chatgpt.com/backend-api/codex/responses (Responses API).

Set ANTHROPIC_MODEL to a model your ChatGPT subscription is allowed to use. Append -fast to a Codex model name to request Codex fast mode for that request without restarting the proxy. For example, gpt-5.5-fast is sent upstream as model gpt-5.5 with service_tier: "priority". An explicit codex.serviceTier / CCP_CODEX_SERVICE_TIER override still takes precedence.

Reasoning effort: Claude Code's output_config.effort value (the one you see in the UI as ◐ medium · /effort) is forwarded as Codex reasoning.effort (low / medium / high / xhigh). Claude Code's max value is sent upstream as xhigh. An explicit codex.effort / CCP_CODEX_EFFORT override still takes precedence and can also force none.

Confirmed working on Plus:

• gpt-5.4
• gpt-5.3-codex

Also verified:

• gpt-5.2
• gpt-5.4-mini

If the resolved model isn't supported by your account, upstream returns a 400 like "The 'gpt-4.1' model is not supported when using Codex with a ChatGPT account.". The proxy surfaces that verbatim.

Auth:

Kimi (Kimi Code)

Upstream: https://api.kimi.com/coding/v1/chat/completions (OpenAI-style chat-completions).

Only one wire model is exposed: kimi-for-coding (its display name in kimi-cli is Kimi-k2.6, 256k context, supports reasoning + image input + video input). kimi-k2.6 and k2.6 are accepted as aliases for the same wire id.

Reasoning effort: Claude Code's output_config.effort value (the one you see in the UI as ◐ medium · /effort) is forwarded as Kimi's reasoning_effort (low / medium / high). Thinking blocks from the upstream model are forwarded to Claude Code and rendered as thinking content. If Claude Code disables thinking, the proxy drops both reasoning_effort and the thinking: {type: "enabled"} flag before forwarding.

Auth:

How it works

sequenceDiagram
    autonumber
    participant CC as Claude Code
    participant P as claude-code-proxy
    participant AUTH as OAuth host<br/>(auth.openai.com or<br/>auth.kimi.com)
    participant U as Upstream API<br/>(chatgpt.com/codex or<br/>api.kimi.com)

    Note over P,AUTH: One-time: PKCE / device OAuth<br/>tokens cached locally for reuse

    CC->>P: POST /v1/messages (Anthropic shape, stream: true)

    alt access token expiring
        P->>AUTH: POST /oauth/token (refresh_token)
        AUTH-->>P: new access (+ rotated refresh)
    end

    P->>P: translate request<br/>• strip Anthropic-only fields<br/>• system blocks → instructions / system message<br/>• tool_use / tool_result ↔ provider-specific shapes<br/>• prompt_cache_key = session id
    P->>U: POST upstream<br/>Bearer + provider-specific headers
    U-->>P: provider SSE<br/>(Codex: output_item.*, output_text.delta, …)<br/>(Kimi: chat.completion.chunk, reasoning_content, …)
    P->>P: reducer: typed events<br/>(thinking / text / tool start/delta/stop, finish)
    P-->>CC: Anthropic SSE<br/>(message_start, content_block_*, message_delta, message_stop)

Commands

serve

Starts the HTTP proxy and blocks. Binds to 127.0.0.1 only. Logs to the platform state directory (rotated at 20 MiB). Set CCP_LOG_STDERR=1 to mirror log lines to stderr while running.

claude-code-proxy serve
PORT=11435 claude-code-proxy serve
CCP_LOG_STDERR=1 claude-code-proxy serve

Prints the supported model → provider mapping on startup. One serve process dispatches to any provider based on the model field in each request. Requests whose model isn't registered with any provider are rejected with HTTP 400 listing the supported ids.

Codex auth commands

codex auth login

Runs the PKCE browser flow against auth.openai.com using the Codex CLI's client ID. Prints a URL, opens a local callback listener on port 1455, waits for the browser to redirect back, and stores the resulting access / refresh tokens in Keychain on macOS or locally on other platforms. The process exits automatically once the tokens are saved.

claude-code-proxy codex auth login

Sign in with your ChatGPT Plus/Pro account, not an OpenAI API account. The token file includes the extracted chatgpt_account_id so the proxy can set the ChatGPT-Account-Id header on every upstream call.

codex auth device

Same OAuth flow, but for headless machines. Prints a short user code and a URL; you enter the code from any browser on any other device, and the CLI polls auth.openai.com until you authorize, then stores the token.

claude-code-proxy codex auth device

Useful over SSH, inside a container, or on any host that can't open a browser.

codex auth status

Shows whether credentials are stored, the account ID, and how long until the access token expires. Non-zero exit if no auth is present.

claude-code-proxy codex auth status

Example output:

Account: 79342a5e-57b7-44ea-bfdc-a83ba070dad6
Expires: 2026-04-28T16:46:04.827Z (in 863946s)
Storage: macOS Keychain

The proxy refreshes the access token 5 minutes before expiry with a single-flight guard, so concurrent requests never trigger stampedes of refresh calls.

codex auth logout

Removes stored auth credentials. On macOS this deletes the Keychain entry. No server call is needed; the refresh token just becomes dead.

claude-code-proxy codex auth logout

Run codex auth login again to re-authenticate.

Kimi auth commands

kimi auth login

Runs a device-code OAuth flow (RFC 8628) against auth.kimi.com using the kimi-cli client ID. Prints a verification URL and a short user code; open the URL in any browser, confirm the code, and the CLI polls until the tokens are issued. Tokens are stored in Keychain on macOS or a mode-0600 file elsewhere.

claude-code-proxy kimi auth login

Sign in with your kimi.com account. The access token has a ~15 minute lifetime; the proxy refreshes it 5 minutes before expiry with a single-flight guard and persists the rotated refresh token.

A persistent device ID is generated on first login next to the Kimi auth file and reused forever — it's bound into the issued JWT, so rotating it would invalidate your token.

kimi auth status

claude-code-proxy kimi auth status

Shows the user ID extracted from the token, expiry time, scope, and storage backend. Non-zero exit if no auth is present.

kimi auth logout

claude-code-proxy kimi auth logout

Removes stored auth credentials (Keychain entry on macOS, file elsewhere). Run kimi auth login again to re-authenticate.

Endpoints

The proxy speaks enough of the Anthropic API for Claude Code:

• POST /v1/messages: the main turn endpoint (streaming and non-streaming)
• POST /v1/messages?beta=true: same (Claude Code always sends ?beta=true)
• POST /v1/messages/count_tokens: local token count via gpt-tokenizer (o200k_base); used by Claude Code's compaction logic
• GET /healthz: liveness check

Configuration

Settings can come from either environment variables or a config.json file. Precedence per setting: env var > config file > built-in default. The config file is optional — env-var-only setups continue to work unchanged.

The file lives at ~/.config/claude-code-proxy/config.json on macOS (deliberately not ~/Library), at %APPDATA%\claude-code-proxy\config.json on Windows, and at ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/config.json on Linux.

{
  "port": 18765,
  "aliasProvider": "codex",
  "codex": {
    "originator": "claude-code-proxy",
    "userAgent": "claude-code-proxy/dev",
    "model": "gpt-5.4",
    "effort": "medium",
    "serviceTier": "fast",
    "baseUrl": "https://chatgpt.com/backend-api/codex/responses",
    "transport": "websocket",
    "previousResponseId": false
  },
  "kimi": {
    "userAgent": "KimiCLI/1.37.0",
    "oauthHost": "https://auth.kimi.com",
    "baseUrl": "https://api.kimi.com/coding/v1"
  },
  "log": {
    "stderr": false,
    "verbose": false
  }
}

A malformed config.json is reported on stderr and ignored; defaults are used in its place. Invalid types for individual keys are warned and skipped without affecting other keys.

Codex uses the WebSocket Responses transport by default. Set CCP_CODEX_TRANSPORT=http to use the older HTTP SSE transport for debugging or compatibility, or CCP_CODEX_TRANSPORT=auto to try WebSocket with HTTP fallback only when setup fails before a request is sent upstream. CCP_CODEX_PREVIOUS_RESPONSE_ID=1 enables opt-in WebSocket continuation for append-only turns. Continuation keeps in-memory state keyed by Claude Code session id, reuses a session WebSocket while it remains open, and sends previous_response_id only when the translated request shape is unchanged and the new input strictly extends the previous transcript. On mismatch, missing state, missing upstream response, closed connections, or setup failure, the proxy clears unsafe continuation state and sends the full request instead. Continuation reduces repeated request upload size, but it does not increase the upstream model context window. Multi-process or load-balanced deployments need sticky sessions or shared state before enabling continuation.

Files

• proxy.log — JSON-lines log, rotated at 20 MiB. It lives at $XDG_STATE_HOME/claude-code-proxy/proxy.log on macOS/Linux and at %LOCALAPPDATA%\claude-code-proxy\proxy.log on Windows (falling back to %USERPROFILE%\AppData\Local). Secrets (authorization, access, refresh, id_token, ChatGPT-Account-Id, …) are redacted before write.
• traffic/ — per-request captures written when CCP_TRAFFIC_LOG=1 is set. Captures live under the state directory, grouped by Claude Code session and request sequence. They include inbound Anthropic requests, translated upstream requests, upstream headers, upstream events, and downstream events. Stream events are written under each request's events/ directory with monotonic sequence numbers so sorted filenames preserve emission order. Token and account headers are redacted, but prompt and tool content are intentionally preserved for debugging.
• config.json — optional configuration file (see table above). It lives at ~/.config/claude-code-proxy/config.json on macOS, ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/config.json on Linux, and %APPDATA%\claude-code-proxy\config.json on Windows.
• Codex tokens — macOS uses Keychain under service claude-code-proxy.codex. Linux uses ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/codex/auth.json. Windows uses %APPDATA%\claude-code-proxy\codex\auth.json.
• Kimi tokens — macOS uses Keychain under service claude-code-proxy.kimi. Linux uses ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/kimi/auth.json. Windows uses %APPDATA%\claude-code-proxy\kimi\auth.json.
• Kimi device ID — persistent UUID bound into the Kimi JWT at login. Linux uses ${XDG_CONFIG_HOME:-$HOME/.config}/claude-code-proxy/kimi/device_id; Windows uses %APPDATA%\claude-code-proxy\kimi\device_id. Reused for the lifetime of the install.

Limitations

• Terms of service: using the Codex or Kimi backends from a non-official client is a gray area. Use at your own risk.
• Rate limits: shared across all clients of your upstream account. Codex's codex.rate_limits.limit_reached and Kimi's HTTP 429 are both surfaced as HTTP 429 with retry-after.
• Codex — image inputs in tool results: Responses API function_call_output only takes a string, so image blocks nested inside tool_result are replaced with a [image omitted: <media_type>] placeholder. Top-level user-message images pass through.
• Kimi — image inputs in tool results: pass through as image_url parts (Kimi accepts them in role:"tool" content).
• Codex — reasoning blocks: not forwarded to Claude Code (dropped), even if the upstream model produced them.
• Kimi — reasoning blocks: forwarded as Anthropic thinking content blocks and rendered by Claude Code. Disable by setting thinking: {"type":"disabled"} in your Anthropic request.
• Session title generation: Claude Code's parallel title-gen request is forwarded upstream like any other structured-output request. This costs a handful of tokens per session rather than being stubbed.
• Codex — output_config.format: translated to Responses API text.format (json_schema with strict: true); other Anthropic-specific output_config fields are dropped.

Development

bunx tsc --noEmit                          # typecheck
bun src/cli.ts serve                       # run locally (routes all providers)
tail -f ~/.local/state/claude-code-proxy/proxy.log | jq .

Install a compiled dev build globally: compile the current working tree to a binary and place it on your PATH without linking:

mkdir -p ~/.local/bin
bun build ./src/cli.ts --compile --outfile ~/.local/bin/claude-code-proxy

Related projects

• claude-history: search Claude Code conversation history from the terminal
• git-surgeon: non-interactive hunk-level git staging for AI agents
• workmux: manage parallel AI coding tasks in separate git worktrees with tmux
• consult-llm: Consult other AI models from your agent workflow