AVLUMO / CONNECT WITHOUT STARTING OVER

Connect without starting over

Documented client entry points, one existing key. Manual guidance is available now; automated setup commands appear only after a signed release is qualified.

No setup release is published here. There is no executable download or invented install command.

Start in the client you already use

One existing key. Choose a client and follow its documented protocol—not a guessed adapter.

Client guides: 16 / 16

Claude Code · Windows

Documented · manual only

Windows: use the client’s documented native route or explicitly supported WSL/remote profile. PowerShell and Bash syntax are not interchangeable; do not weaken execution policy. Use the installed client’s release-specific Windows/WSL or macOS/Linux route; inspect the actual user/editor/remote profile.

MODEL_ID is a placeholder. Select an exact currently published model allowed for your access; no model is recommended or invented here.

Connection field map · not a command

PROTOCOL     anthropic_messages
BASE_URL     https://api.avlumo.com
REQUEST_PATH /v1/messages
MODEL_ID     MODEL_ID

Planned Avlumo origin, not live endpoint proof. Confirm the field’s suffix behavior in the manual.

Enter the key in your client, not here

Use its local masked prompt or private API Key field. This page never accepts, transmits or stores a key.

Complete client manual

Claude Code — manual setup

Manual guidance, documentation reviewed on 2026-10-03. No Avlumo client/model/OS integration is qualified; native T02/T03 have not passed. The URLs below are planned Avlumo origins, not tested live endpoints. The helper does not install clients, accept a key, fetch a live model catalog, edit installed-client configuration, or run an inference test. Its separate Linux synthetic preview/apply/restore workflow is available through guide workflow, not installed-client proof.

MODEL_ID is a placeholder, not a model recommendation. Replace it only with a currently published model allowed for your access and separately qualified for this client/protocol. If that information is missing, stop before any billed request. Reuse your existing key; setup does not change your access, purchased allowance, pricing profile, or source policy. No new account or purchase is required for these instructions.

1. Prerequisites and OS route

Use an already installed Claude Code CLI release supported by its own OS instructions: shell on macOS/Linux; PowerShell for native Windows, or a shell inside WSL. Meet the prerequisites of that exact release rather than assuming feature parity. The helper does not install Claude Code or alter OS execution controls.[4]

2. Where to configure and what to preserve

Use a separate terminal session for the initial setup. The documented persistent CLI settings are ~/.claude/settings.json (Windows %USERPROFILE%\.claude\settings.json) with an env block; managed settings can override local values. A project file is not a safe default for a credential. The CLI, editor extension and desktop app have different configuration paths; this guide targets the CLI only.[23]

Before manual changes, use your normal backup tool to save affected configuration and credential-store files in a durable private location outside any helper temporary directory. Protect secret-bearing backups: preserve owner-only POSIX permissions or restrictive Windows ACLs; this helper does not establish or verify those ACLs. Record only the nonsecret fields you changed. Do not replace a whole file or remove other providers, MCP servers, plugins, skills, hooks, permissions, or credentials. If your format differs from the example, use that release's supported UI/commands and primary documentation; do not guess a parser or apply this fragment blindly.

3. Protocol, base URL and model

The planned base is https://api.avlumo.com **without** /v1: Claude Code appends /v1/messages. This requires Anthropic Messages with the expected headers and complete streaming/tool semantics, not merely an OpenAI endpoint. Use the published compatible MODEL_ID explicitly. Advanced MCP tool discovery, model aliases and small/background model selection need separate qualification; do not silently map them to an arbitrary model.[24]

{"env":{"ANTHROPIC_BASE_URL":"https://api.avlumo.com"}}

The fragment is nonsecret and illustrative; merge only its owned field into the correct scope, never overwrite the existing env block.

4. Enter the key locally

For a bearer gateway credential, the documented variable is ANTHROPIC_AUTH_TOKEN; ANTHROPIC_API_KEY instead uses x-api-key. Do not set both blindly or delete an existing login/helper. Confirm the selected Avlumo authentication contract before sending it.[23] In a dedicated **Bash** session (use bash if your shell is zsh), or Windows PowerShell, masked local input avoids key literals in command history/argv:

export ANTHROPIC_BASE_URL='https://api.avlumo.com'
read -r -s -p 'Avlumo API key: ' ANTHROPIC_AUTH_TOKEN
printf '\n'
export ANTHROPIC_AUTH_TOKEN
$env:ANTHROPIC_BASE_URL = 'https://api.avlumo.com'
$localKey = Read-Host 'Avlumo API key' -AsSecureString
$env:ANTHROPIC_AUTH_TOKEN = [System.Net.NetworkCredential]::new('', $localKey).Password
$localKey = $null

Only this process environment and children receive the value; do not write it into global OS variables, browser storage or a tracked file. Close the dedicated terminal to discard this temporary configuration. For persistence, use a qualified client credential helper/private user scope, keeping the key out of copied commands.

5. Verification: free versus billed

Free local checks: read this guide, inspect nonsecret settings and the configured model/provider selection. avlumo-setup doctor only inventories the local machine; it does not authenticate your key or prove network/model/tool compatibility. Client startup, model discovery and a button named Verify/Test connection are not automatically free or offline. Authenticated read-only discovery can be called free only after the exact Avlumo route and its billing are qualified; this increment makes no such request.

An optional real prompt, tool cycle or streaming test is billed and can invoke background models or retries. Run it only after access is active, native gates and client/model compatibility are passed, you deliberately accept the relevant tariff and a bounded budget, and the selected model is published. Use a disposable nonsecret workspace, retain existing approval restrictions, and compare the one purchased allowance and actual usage afterward. Do not send private project contents or keys in a test prompt. No paid test has been executed here.

After deliberately authorizing any client traffic, claude --model MODEL_ID starts the selected session and /status shows the configured route and credential source; it is a local configuration check, not proof of billed inference. Never attach a debug log containing credentials. A later real prompt is the separately billed test.[23]

claude --model MODEL_ID
/status
6. Undo, errors and remaining limits

Stop the affected client session before undoing your edits. Restore the previous selection and remove only the Avlumo fields/credential you added through the client-supported mechanism. Compare the current file with your private backup and change record: if you or the client edited it later, merge the intended reversal instead of overwriting the current file. Keep unrelated edits, providers, credentials, MCP/plugins/skills/hooks and permissions. Do not delete an entire configuration directory, reset the client, revoke your shared Avlumo key merely to undo a local setting, or restore a full old file over later changes. Automatic installed-client backup/restore is not enabled. The Linux synthetic workflow is separate; see guide workflow.

For 401, check credential source and original-access state locally without sharing the key. For an unavailable model or protocol, check the actual published compatibility information; do not guess another alias. Activation-pending stock does not permit inference. Insufficient allowance or source outage does not authorize direct checkout, account conversion, or another support channel. Use the support/refill action actually permitted by your original access; share only a sanitized reference and client version. Streaming, tool calls, context limits, retries, background models, hosted tools and billing remain unqualified until separately tested.

Sources

[4] https://code.claude.com/docs/en/setup [23] https://code.claude.com/docs/en/llm-gateway-connect [24] https://code.claude.com/docs/en/llm-gateway-protocol

Reusable protocol guide

Reusable protocol guide

Documentation/configuration evidence is separate from actual Avlumo publication, local installed-version testing and billed compatibility. These are planned Avlumo request shapes, not live API claims. MODEL_ID is non-executable data from your access’s published catalog.

Chat Completions

For SDK-style versioned Base URL fields, use https://api.avlumo.com/v1; the client appends /chat/completions. For a separate Host + API Path form (Chatbox), use host https://api.avlumo.com and path /v1/chat/completions. Cherry Studio’s documented ordinary API Address auto-appends the full versioned path, so it uses the root. Do not generalize one field’s suffix behavior to another app.

PROTOCOL      chat_completions
VERSION_BASE  https://api.avlumo.com/v1
REQUEST_PATH  /v1/chat/completions
AUTH_HEADER   Authorization: Bearer <local client credential>
MODEL_ID      MODEL_ID

Chat support alone does not prove native tools, streaming completion, vision, reasoning parameters, embedding, RAG, audio or automatic retries. Configure only the capability explicitly published for that model and client.

Responses

For SDK-style base fields, use https://api.avlumo.com/v1; the client appends /responses. Codex’s custom provider requires wire_api responses. Continue’s default transport can switch by model family: useResponsesApi false explicitly retains Chat Completions. Open WebUI’s Responses setting is experimental; Cherry Studio exposes Responses in newer endpoint dialogs. Version, endpoint format and actual model capability require separate qualification.

PROTOCOL      responses
VERSION_BASE  https://api.avlumo.com/v1
REQUEST_PATH  /v1/responses
AUTH_HEADER   Authorization: Bearer <local client credential>
MODEL_ID      MODEL_ID

A Responses endpoint does not imply server-side persistence, previous_response_id, hosted tools, background jobs or WebSockets. Keep stateless/ordinary HTTP behavior unless the exact deployed route advertises and qualifies more. Do not silently map Responses-only clients onto chat.

Anthropic Messages

Anthropic SDK-style root fields and Claude Code use https://api.avlumo.com without /v1, because their client appends /v1/messages. A raw request path is /v1/messages. A provider exposing a Claude-named model through OpenAI chat is not necessarily a Messages endpoint. Claude Code bearer-token selection must follow its canonical original manual rather than blindly setting both API-key and auth-token variables.

PROTOCOL      anthropic_messages
SDK_ROOT      https://api.avlumo.com
REQUEST_PATH  /v1/messages
AUTH_HEADER   x-api-key: <local client credential>
VERSION       anthropic-version: 2023-06-01
MODEL_ID      MODEL_ID

Check native message/tool-result block shapes and the complete SSE start/delta/stop sequence. Images, extended thinking, prompt caching, tool-stream fragments and minimum charges are model/route-specific. Do not copy supplier prices, free labels, auth prefixes or routing claims into Avlumo retail promises.

Evidence and field rules

The current primary client docs specify these configuration routes. InferHub’s reviewed public schema 1.0.0 declares /v1/chat/completions, /v1/responses, /v1/messages and /v1/models; its main docs introduce chat and Messages first. OpenRouter’s quickstart distinguishes its own /api/v1 base. Those supplier/competitor examples establish useful documentation organization, not Avlumo authentication, cost, model admission or live compatibility.

Canonical command/configuration blocks remain identical in EN/RU. Do not substitute credentials, arbitrary model strings or downloaded catalog content into executable commands. UI key entry and private configuration stay local to the chosen client. Unknown fields/formats stay manual without automatic mutation.

Sources

[1] https://docs.continue.dev/customize/model-providers/top-level/openai [6] https://inferhub.dev/docs [7] https://inferhub.dev/api/openapi.json [8] https://openrouter.ai/docs/quickstart [13] https://docs.openwebui.com/getting-started/quick-start/connect-a-provider/starting-with-open-responses [17] https://docs.chatboxai.app/en/guides/providers [28] https://docs.cherryai.com.cn/pre-basic/providers/providers