Klide AI · Docs

Connect models to your workflow.

Klide AI speaks the Anthropic Messages API. Anything that works with the official API works here once you change two values: the base URL and the API key. You'll find both on your Console page.

1. 01Create a key
2. 02Pick a protocol
3. 03Connect your tool
4. 04Verify

Video tutorialSign up, add credit, create a key and chat with any model

Quick start

In every snippet on this page, {base_url} means that full address and {api_key} means your key. Replace them with your own values, and never paste a real key into a page, chat or screenshot you share.

Claude DesktopDeveloper Mode → Configure Third-Party Inference

The desktop app ignores environment variables and settings.json. It has its own screen for pointing at a custom API address.

1. Update Claude Desktop to the latest version, then open it.

2. Turn on Developer Mode. On macOS use the menu bar. On Windows open the application menu ☰ at the top left of the sign-in screen.

   Help → Troubleshooting → Enable Developer Mode

3. A Developer menu now exists. Open Developer → Configure Third-Party Inference…

4. Set Inference provider to Gateway, then fill the Gateway credentials card:

   Field	Enter
   Gateway base URL	{base_url} No trailing slash and no /v1; the app adds those itself.
   Gateway API key	Your key: {api_key}
   Credential kind	Static API key
   Gateway auth scheme	Bearer or x-api-key. Klide AI accepts both.

5. Leave Models empty. The app asks Klide AI for its list and shows every Claude ID your key can use. To lock it to a few, list full IDs instead, for example claude-opus-5-5 and claude-sonnet-4-6. Give a [1m] ID the supports1m flag so the app knows it has the big window.

6. Click Apply Changes. The app relaunches using Klide AI. Start a chat and send a short message to confirm.

VS CodeThe Claude Code extension

1. Install Claude Code (publisher: Anthropic) from the Extensions panel.

2. Press Ctrl+Shift+P (Cmd+Shift+P on macOS) and run Preferences: Open User Settings (JSON).

3. Add this block inside the top-level braces. If the file already has entries, add a comma after the previous one.

   "claudeCode.environmentVariables": [
     { "name": "ANTHROPIC_BASE_URL", "value": "{base_url}" },
     { "name": "ANTHROPIC_AUTH_TOKEN", "value": "{api_key}" }
   ]

4. Run Developer: Reload Window, open the Claude panel and type /status. You should see an Anthropic base URL line with your address. If you see a claude.ai login instead, the setting wasn't picked up.

Claude Code CLIWindows, macOS and Linux · also covers the JetBrains plugin

Option 1: settings file (set once, works everywhere)

Add an env block to ~/.claude/settings.json (on Windows: %USERPROFILE%\.claude\settings.json). It applies to the terminal, VS Code and JetBrains, and needs no restart.

{
  "env": {
    "ANTHROPIC_BASE_URL": "{base_url}",
    "ANTHROPIC_AUTH_TOKEN": "{api_key}"
  }
}

The file must be strict JSON: no comments, no trailing commas.

Option 2: environment variables for one session

# macOS / Linux
export ANTHROPIC_BASE_URL="{base_url}"
export ANTHROPIC_AUTH_TOKEN="{api_key}"
claude

# Windows PowerShell
$env:ANTHROPIC_BASE_URL = "{base_url}"
$env:ANTHROPIC_AUTH_TOKEN = "{api_key}"
claude

PowerShell's $env: lasts only for that window. To keep it, use Option 1, or run setx and then open a new terminal (an already-open one never sees it).

Check it worked

Start claude and run /status. You want an Anthropic base URL line and an Auth token line. If it shows a Login method with a claude.ai account instead, run /logout and start again. Use only one credential variable: if you also have ANTHROPIC_API_KEY set, remove it, or you'll see an "Auth conflict" warning.

Codex CLIThe OpenAI terminal agent, running Claude models

Codex speaks the OpenAI Responses API, which Klide AI translates to Claude and back. It also needs a short model list, so it knows Claude can edit files with its patch tool. Without that list, Codex falls back to shell commands for edits, and those fail on Windows.

1. Download the model list

# macOS / Linux
mkdir -p ~/.codex
curl -fsSL -o ~/.codex/klide-models.json https://api.klideai.store/assets/codex-models.json

# Windows PowerShell
New-Item -ItemType Directory -Force "$env:USERPROFILE\.codex" | Out-Null
Invoke-WebRequest https://api.klideai.store/assets/codex-models.json -OutFile "$env:USERPROFILE\.codex\klide-models.json"

2. Add Klide AI to config.toml

Open ~/.codex/config.toml (on Windows: %USERPROFILE%\.codex\config.toml), creating it if it doesn't exist. Put the first three lines at the very top, above any [section], and write the full path of the file you downloaded. On Windows use forward slashes, like C:/Users/you/.codex/klide-models.json.

model = "claude-sonnet-4-6"
model_provider = "klide"
model_catalog_json = "/Users/you/.codex/klide-models.json"

[model_providers.klide]
name = "Klide AI"
base_url = "{base_url}/v1"
env_key = "KLIDE_API_KEY"
wire_api = "responses"

3. Set your key and start it

# macOS / Linux
export KLIDE_API_KEY="{api_key}"
codex

# Windows PowerShell
$env:KLIDE_API_KEY = "{api_key}"
codex

Switch models with /model: claude-opus-5-5, claude-sonnet-4-6 and claude-haiku-4-5 are on the list. If Codex warns Model metadata … not found, the model_catalog_json path is wrong or the model isn't on the list. Codex's built-in web search only works with OpenAI's own servers, so it is left out here.

API & SDKs

curl

curl {base_url}/v1/messages \
  -H "x-api-key: {api_key}" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model": "claude-sonnet-4-6", "max_tokens": 512,
       "messages": [{"role": "user", "content": "Hello"}]}'

Python and TypeScript

# Python
from anthropic import Anthropic
client = Anthropic(base_url="{base_url}", api_key="{api_key}")

// TypeScript
import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic({ baseURL: "{base_url}", apiKey: "{api_key}" });

OpenAI SDKs and Cursor OpenAI format

Klide AI also answers the OpenAI Chat Completions format, so apps that only know the OpenAI API can use your key. Requests are translated to the Claude API and back, and you still pick a Claude model by its ID. Two differences from the Claude setup above: the base URL ends in /v1, and the key goes in as an OpenAI key.

# Python
from openai import OpenAI
client = OpenAI(base_url="{base_url}/v1", api_key="{api_key}")
reply = client.chat.completions.create(
    model="claude-sonnet-4-6",
    messages=[{"role": "user", "content": "Hello"}],
)

// TypeScript
import OpenAI from "openai";
const client = new OpenAI({ baseURL: "{base_url}/v1", apiKey: "{api_key}" });

# curl
curl {base_url}/v1/chat/completions \
  -H "Authorization: Bearer {api_key}" \
  -H "content-type: application/json" \
  -d '{"model": "claude-sonnet-4-6", "messages": [{"role": "user", "content": "Hello"}]}'

The newer Responses API works with the same client:

# Python
reply = client.responses.create(model="claude-sonnet-4-6", input="Hello")
print(reply.output_text)

# curl
curl {base_url}/v1/responses \
  -H "Authorization: Bearer {api_key}" \
  -H "content-type: application/json" \
  -d '{"model": "claude-sonnet-4-6", "input": "Hello"}'

Cursor: open Settings → Models, turn on Override OpenAI Base URL and enter {base_url}/v1. Paste your key into the OpenAI API Key box. Then Add Model, type a Claude ID such as claude-sonnet-4-6, and use Verify. Cursor sends these requests from its own servers, and which of its modes accept a custom model depends on your Cursor version. If it refuses a Claude model name, tell us what it says.

What carries over: text, images, system messages, tool calling, streaming, max_tokens and stop. The Responses API adds PDFs, freeform (custom) tools, previous_response_id, and retrieving or deleting a stored response. Temperature is capped at 1, and top_p is only used when you leave temperature out.

JSON mode (response_format, or text.format in the Responses API) works by instruction: the model is given your schema and told to reply with JSON only. Nothing enforces the schema, so validate what you parse.

Not supported: more than one choice (n), logprobs, embeddings, audio, image generation, built-in tools that run on OpenAI's side (web search, file search, code interpreter), background mode, and the Assistants, Files and Batch APIs.

Prompt caching

Works as on the Anthropic API: add "cache_control": {"type": "ephemeral"} to a system or content block on /v1/messages. Claude Code, Cline and the other Anthropic-format tools already do this for you. When a later request starts with the same text, the repeated part is read from cache and the response says how much in usage.cache_read_input_tokens. In OpenAI format the same number is usage.prompt_tokens_details.cached_tokens.

Caching is best effort: a repeat can miss, so check cache_read_input_tokens rather than assume. It only applies to long, stable prefixes of a few thousand tokens. Put the part that never changes at the start of the prompt and the part that does at the end.

What works and what doesn't

Models

The full list, with copy buttons, is on the home page. To get the exact list your key can use, ask the API:

curl {base_url}/v1/models -H "x-api-key: {api_key}"

• Pick an ID, don't type a nickname. sonnet or opus alone won't resolve. Use a full ID such as claude-sonnet-4-6.
• Only IDs on the list work. Anything else, including a spelling variant, is refused with a 404. Opus 5.5 is listed both ways: claude-opus-5-5 and claude-opus-5.5 are the same model.
• [1m] means a 1M-token context window. Quote it in the shell: "claude-opus-5-5[1m]".
• Dated IDs are snapshots. claude-haiku-4-5-20251001 never changes; claude-haiku-4-5 follows the latest.
• Older generations (3.x, 4.0, 4.1) are listed for compatibility. New work should use the newer IDs.

Limits & usage

These are the defaults; your key may be set differently. The 429 message names the exact limit you reached. Need more? Ask on Telegram. One key shared across devices or teammates shares one balance and one rate limit.

Open Console, paste your base URL and API key, and you'll see your remaining balance, requests and tokens over the last 24 hours, latency, and a per-model breakdown. Input and output tokens both count against your balance.

Privacy & securityWhat we keep, and what we don't

Your data is never stored by Klide AI. Your prompts, files and model replies pass through the gateway and are not saved.

One exception, for the OpenAI Responses API only: a request sent with store left on (the API's default) is held in memory for up to an hour, so a follow-up can continue it with previous_response_id. It is never written to disk, only the key that made it can read it, and you can delete it early. Send "store": false to skip this. Codex CLI already does.

• Your keys are encrypted at rest. Only you can see them, on your Account page after signing in.
• Your key stays private. It is never shown to anyone else.
• All traffic is HTTPS.
• Signing in keeps you signed in for 30 days on that browser. Press Sign out on a shared computer.

Keep your key secret like a password: don't commit it, don't paste it in screenshots, and never put it in a shared .vscode/settings.json or .claude/settings.json. If it leaks, message us and we'll disable it.

Troubleshooting

FAQ

Getting started

Usage & limits

Billing & expiry

Privacy & security

Still stuck? Message support on Telegram. Include the error text and which tool you're using.