Docs

Point your toolsat LotusAI.

Two environment variables and you are done. Everything that speaks the Anthropic or OpenAI API keeps working; only the base URL changes.

Quick start API reference Check a key

01 / Quick start

Two variables. That is the setup.

Every Anthropic-format tool reads the same two environment variables. Set them and the tool talks to LotusAI.

macOS / Linux (bash, zsh)

bash

export ANTHROPIC_BASE_URL=https://api.lotusai.pro
export ANTHROPIC_AUTH_TOKEN=sk-lotus-...

Windows (PowerShell)

powershell

$env:ANTHROPIC_BASE_URL = "https://api.lotusai.pro"
$env:ANTHROPIC_AUTH_TOKEN = "sk-lotus-..."

Get a key

Keys are issued by authorised partners. They start with sk-lotus-.

Set the variables

Base URL plus token, in your shell or the tool config.

Keep building

Streaming, tool use, vision and long context all pass straight through.

Base URL: https://api.lotusai.pro. Auth: send the key as x-api-key or Authorization: Bearer; both work on every endpoint. For OpenAI-format tools, add /v1 to the base URL.

02 / One-line installer

Or let a script do it.

The installer asks for your key, writes the two variables to your shell profile (or your Windows user environment), configures Claude Code, installs it via npm if it is missing, and verifies the key. Re-run it any time to change keys.

macOS / Linux

bash

curl -fsSL https://lotusai.pro/setup.sh | bash

Windows (PowerShell)

powershell

irm https://lotusai.pro/setup.ps1 | iex

Idempotent: a marked block in your profile is replaced, never appended twice.
Pass the key non-interactively with LOTUSAI_API_KEY=sk-lotus-... in the environment.
Writes ~/.claude/settings.json so Claude Code works from any terminal or editor.
Prefer to read before you run? Both scripts are plain text at the URLs above.

03 / Set up your tool

Pick your editor.

Manual configuration for each tool. Replace sk-lotus-... with your key.

Claude Code

Anthropic format

1Run the one-line installer above, or set the two environment variables by hand.
2Or put them in ~/.claude/settings.json so they apply to every project.
3Restart Claude Code. Run /status inside it to confirm the base URL.

shell

bash

export ANTHROPIC_BASE_URL=https://api.lotusai.pro
export ANTHROPIC_AUTH_TOKEN=sk-lotus-...

~/.claude/settings.json

json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.lotusai.pro",
    "ANTHROPIC_AUTH_TOKEN": "sk-lotus-..."
  }
}

Optional: pin a model with ANTHROPIC_MODEL, for example claude-fable-5-1[1m]. The [1m] suffix only tells Claude Code to size its own context budget at 1M; the API ignores it.

04 / API reference

Two formats, one key.

Anthropic Messages at /v1/messages, OpenAI Chat Completions at /v1/chat/completions. Same key, same window, same models.

POST/v1/messagesAPI key

Anthropic Messages format, request and response. Supports system, multi-turn messages, image blocks, tools with tool_use and tool_result, and stream: true for server-sent events. Keys are sent as x-api-key; the anthropic-version header is accepted and optional.

basic request

bash

curl https://api.lotusai.pro/v1/messages \
  -H "x-api-key: sk-lotus-..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Explain this regex: ^sk-lotus-" }]
  }'

streaming

bash

curl -N https://api.lotusai.pro/v1/messages \
  -H "x-api-key: sk-lotus-..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "max_tokens": 1024,
    "stream": true,
    "messages": [{ "role": "user", "content": "Stream me a limerick." }]
  }'

# event: message_start
# event: content_block_start
# event: content_block_delta   (repeats)
# event: content_block_stop
# event: message_delta
# event: message_stop

tool use (request body)

json

{
  "model": "claude-sonnet-5",
  "max_tokens": 1024,
  "tools": [{
    "name": "get_weather",
    "description": "Current weather for a city",
    "input_schema": {
      "type": "object",
      "properties": { "city": { "type": "string" } },
      "required": ["city"]
    }
  }],
  "messages": [{ "role": "user", "content": "Weather in Lisbon?" }]
}

// The reply carries a tool_use block:
// { "type": "tool_use", "id": "toolu_...", "name": "get_weather", "input": { "city": "Lisbon" } }
// Run the tool, then send a tool_result block with the same id in the next user turn.

POST /v1/messages/count_tokens is also available and takes the same body, returning { "input_tokens": n } without sending the request.

POST/v1/chat/completionsAPI key

OpenAI Chat Completions format. The request is translated to the Messages API and the reply is translated back, so messages with system roles, tools with tool_calls, and stream: true all behave as the OpenAI SDK expects. Use Authorization: Bearer for the key.

chat completions

bash

curl https://api.lotusai.pro/v1/chat/completions \
  -H "Authorization: Bearer sk-lotus-..." \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "stream": false,
    "messages": [
      { "role": "system", "content": "You are a terse senior engineer." },
      { "role": "user", "content": "Why is my build slow?" }
    ]
  }'

GET/v1/modelsNo auth

Lists every model id with its display_name and context_window. Public, so tools can populate their model pickers before a key is entered. GET /v1/models/:id returns one entry.

response (truncated)

json

{
  "data": [
    {
      "id": "claude-fable-5-1",
      "display_name": "Claude Fable 5.1",
      "type": "model",
      "context_window": 1000000
    },
    {
      "id": "claude-fable-5",
      "display_name": "Claude Fable 5",
      "type": "model",
      "context_window": 1000000
    },
    {
      "id": "claude-opus-5",
      "display_name": "Claude Opus 5",
      "type": "model",
      "context_window": 1000000
    }
  ],
  "has_more": false
}

GET/api/key-status?key=sk-lotus-...No auth

Everything the Check key page shows, as JSON: plan, active and expiry flags, window usage as a percentage, per-family split, reset time, request counts and the last twenty requests. Token budgets are never exposed as raw numbers. An unknown key returns { "status": "error", "error": "API key not found" }.

response

json

{
  "status": "found",
  "name": "Team key",
  "keyPrefix": "sk-lotus-a1b2",
  "isActive": true,
  "planName": "Studio",
  "rateLimit": 60,
  "unlimited": false,
  "usagePercent": 42.5,
  "byModel": { "fable": { "percent": 10.2 }, "opus": { "percent": 0 }, "sonnet": { "percent": 30.1 }, "haiku": { "percent": 2.2 } },
  "windowActive": true,
  "windowStartedAt": "2026-10-03T08:10:00.000Z",
  "windowResetAt": "2026-10-03T13:10:00.000Z",
  "expiresAt": "2026-11-01T00:00:00.000Z",
  "lastUsedAt": "2026-10-03T09:42:11.000Z",
  "createdAt": "2026-09-01T12:00:00.000Z",
  "isExpired": false,
  "isOverLimit": false,
  "totalRequests": 1842,
  "last24h": { "requests": 96 },
  "recentLogs": [{ "model": "claude-sonnet-5", "status": 200, "time": "2026-10-03T09:42:11.000Z" }]
}

05 / Models

Every model, one key.

Pass the id in the model field. This list is built from GET /v1/models and refreshes every five minutes.

Model	Best for	Context
Fable 5.1new `claude-fable-5-1` The most capable model on the list: demanding reasoning and long-horizon agentic work. Knows the world up to June 2026, the most recent cutoff of anything we serve.	The most capable model on the list: demanding reasoning and long-horizon agentic work. Knows the world up to June 2026, the most recent cutoff of anything we serve.	1M
Fable 5 `claude-fable-5` The Fable before 5.1. Still holds its own on long agentic runs and whole-codebase work. Keep it pinned if your prompts are tuned to it.	The Fable before 5.1. Still holds its own on long agentic runs and whole-codebase work. Keep it pinned if your prompts are tuned to it.	1M
Opus 5new `claude-opus-5` The newest Opus. Thinks before it answers by default, and knows the world up to May 2026. Point it at the hard stuff.	The newest Opus. Thinks before it answers by default, and knows the world up to May 2026. Point it at the hard stuff.	1M
Sonnet 5new `claude-sonnet-5` Frontier-class intelligence at Sonnet speed. The new default for everyday building.	Frontier-class intelligence at Sonnet speed. The new default for everyday building.	1M
Opus 4.8 `claude-opus-4-8` Adaptive thinking and agentic coding. Whole-repo refactors that need it to actually hold the plot.	Adaptive thinking and agentic coding. Whole-repo refactors that need it to actually hold the plot.	1M
Opus 4.7 `claude-opus-4-7` Still a monster. Keep pinning it if your prompts are tuned to it.	Still a monster. Keep pinning it if your prompts are tuned to it.	1M
Opus 4.6 `claude-opus-4-6` Deep reasoning on big, gnarly problems.	Deep reasoning on big, gnarly problems.	1M
Sonnet 4.6 `claude-sonnet-4-6` The everyday sweet spot: most of what you build, most days.	The everyday sweet spot: most of what you build, most days.	1M
Opus 4.5 `claude-opus-4-5` A known quantity for pipelines you don’t want to re-tune.	A known quantity for pipelines you don’t want to re-tune.	200K
Sonnet 4.5 `claude-sonnet-4-5-20250929` Solid general-purpose work when you want the older behaviour.	Solid general-purpose work when you want the older behaviour.	200K
Haiku 4.5 `claude-haiku-4-5-20251001` Autocomplete, quick edits, classification, anything high-volume where latency is the whole game.	Autocomplete, quick edits, classification, anything high-volume where latency is the whole game.	200K
Opus 4.1 `claude-opus-4-1-20250805` Pinned for reproducibility.	Pinned for reproducibility.	200K
Opus 4 `claude-opus-4-20250514` Pinned for reproducibility.	Pinned for reproducibility.	200K
Sonnet 4 `claude-sonnet-4-20250514` Pinned for reproducibility.	Pinned for reproducibility.	200K

06 / The 5-hour window

Usage that refills.

Every key has a usage budget per five-hour window instead of a monthly meter. Use it, let it refill, use it again.

How the clock works

Starts on first use. The window opens with the first request that consumes tokens, not at a fixed hour.
Runs for five hours. Every request in that span counts towards the same budget.
Then resets completely. Usage goes back to zero and the next request opens a new window.

What counts

Input and output tokens of each request, weighted per model family. Prompt-cache reads do not count. Larger models draw the budget down faster than smaller ones.

What happens at the limit

Requests are answered with a short message stating the time until the window refills, so your tool keeps running instead of crashing on an error. Requests resume the moment the window resets.

Where to watch it

Check key shows the gauge, the per-model split and a live countdown. The same data is at GET /api/key-status.

07 / Rate limits and status codes

What the gateway returns.

Errors use the Anthropic error envelope on /v1/messages and the OpenAI envelope on /v1/chat/completions.

Per key

req / min

Each key has a requests-per-minute ceiling set by the partner. It is shown on the Check key page. Exceeding it returns 429.

Per IP

200 / min

A global guard against floods, counted per source address across all endpoints. Normal editor use never approaches it.

Per window

5 hours

The token budget described above. Reaching it does not return an error; see the window section.

Code	error.type	Meaning
200		Success. Also used for the limit-reached, expired and disabled notices, which arrive as a normal assistant message so tools display them.
400	invalid_request_error	Malformed body, empty messages array, unknown parameter, or a request too large to process.
401	authentication_error	Missing or unrecognised key.
404	not_found_error	Unknown model id or path.
413	invalid_request_error	Payload exceeds the size limit. Shrink images or trim history.
429	rate_limit_error	Per-key or per-IP request rate exceeded. Honour the Retry-After header and back off.
500 / 502	api_error	Something failed on our side or between us and the model. Retry with backoff.
503	maintenance	Planned maintenance, with Retry-After in seconds.
529	overloaded_error	Model capacity is saturated. Retry after a short pause.

error envelope

json

{
  "type": "error",
  "error": {
    "type": "rate_limit_error",
    "message": "Rate limit reached. Please try again shortly."
  }
}

08 / Troubleshooting

When something looks off.

The short list. If yours is not here, Telegram is the fastest route to a human.

401 authentication_error+

The key was not received or was not recognised. Check for a trailing space, make sure it begins with sk-lotus-, and confirm it is sent as x-api-key or Authorization: Bearer. Then paste it into Check key.

The reply says the usage limit was reached+

Your 5-hour window is full. The reply includes the time until it refills, and the countdown is also on the Check key page. Nothing to fix; it resets on its own.

Model not found+

Pass an exact id from the models list, for example claude-sonnet-5. GET /v1/models returns the live list.

Cursor says the endpoint failed verification+

Cursor speaks OpenAI format, so the base URL needs the /v1 suffix: https://api.lotusai.pro/v1. Add the model id as a custom model after verifying.

Claude Code still goes to the default endpoint+

Environment variables set in one terminal do not reach a Claude Code launched elsewhere. Put them in ~/.claude/settings.json (the installer does this), then restart Claude Code and run /status.

Context window exceeded+

Long sessions can outgrow a model's window. The gateway compacts automatically and retries; if it still fails, start a fresh conversation or switch to a model with a larger context.

Changes to a config file do nothing+

Almost every tool reads its config at launch. Restart the editor or CLI after editing.

503 with Retry-After+

Planned maintenance. Wait the number of seconds in the header and retry. Live health is on the status page.

Telegram support [email protected]

Ready

Have a key? Go build.

Still need one? Keys are issued by authorised partners, and we will put you in touch with one.

Get access Read the FAQ