LibreChat with the AI Stupid Level Smart Router

A custom endpoint in librechat.yaml puts the routing modes in LibreChat’s model menu, with fetch: true to list every model the router offers and auto-fastest-quality for conversation titles.

Routing on benchmark results

A routing mode as the model name sends each request to the model placing highest on the relevant board — coding, reasoning or tool use — ranked on the average of its last five benchmark runs, never a single run. Statistically tied leaders are split by your tie-break preference, the cheaper measured cost per task by default. Drift-aware routing moves a model whose drift detector has fired for that skill behind every unflagged model.

Your own provider keys

You pay OpenAI, Anthropic, Google, DeepSeek, Moonshot AI (Kimi) or Z.ai (GLM) directly for tokens. Provider keys are encrypted at rest (AES-256-GCM) and never shown again.

No token markup

The router adds nothing to token prices. Free includes 1,000 routed requests a month, and only successful requests count.

The model that answered, every time

Every response carries X-AISM-Model and X-AISM-Reasoning: which model answered and why, including any fallback. If the chosen model fails before answering, the router tries your own fallback order, then up to 2 automatic fallbacks from different providers.

1. 01

   Create a router key

   Go to API keys, click Create key and copy it somewhere safe. The key starts with aism_ and is the one key for every tool you connect.

2. 02

   Add your provider keys

   Under Providers, add keys for OpenAI, Anthropic, Google, DeepSeek, Kimi or GLM. The more providers you add, the better the routing. You pay them directly for tokens; the router adds nothing to token prices, and provider keys are encrypted at rest (AES-256-GCM) and never shown again.

3. 03

   Put the key in LibreChat’s environment

   The sample reads the key as ${ASL_ROUTER_KEY}, so set that variable where LibreChat runs (its .env, or the container’s environment) to your aism_ key.

   .envShell

   ASL_ROUTER_KEY=aism_your_key_here

4. 04

   Add the custom endpoint to librechat.yaml

   One entry under endpoints.custom: the base URL, the key from the environment, the routing modes as default models, fetch: true to list every model the router offers, and auto-fastest-quality as the title model.

   librechat.yamlYAML

   endpoints:
     custom:
       - name: "AI Stupid Level"
         apiKey: "${ASL_ROUTER_KEY}"
         baseURL: "https://aistupidlevel.info/v1"
         models:
           default: ["auto-coding", "auto-reasoning", "auto-tooling", "auto-best", "auto-value"]
           fetch: true
         titleConvo: true
         titleModel: "auto-fastest-quality"
         modelDisplayLabel: "AI Stupid Level"

5. 05

   Restart LibreChat and pick AI Stupid Level

   LibreChat reads librechat.yaml when it starts. The endpoint then appears in the menu as AI Stupid Level with the routing modes as models, and every reply is routed by the model name you choose.

auto-best

Overall. Best combined score regardless of category; among statistically tied leaders, your tie-break decides (the cheaper by default).Top of the Combined board in the last 24 hours: DeepSeek V4.1 Flash, in a statistical tie for first place, of 22 ranked models.

auto-value

Value. Most combined-score points per measured dollar, among models within 5 points of the best one you can use.

auto-reasoning

Reasoning. Best for complex analysis, math, logic.Top of the Reasoning board in the last 24 hours: Claude Fable 5.1, in a statistical tie for first place, of 22 ranked models.

auto-fastest-quality

Speed. Among models within 5 points of the best one you can use, the fastest (measured latency) — auto-fastest has no quality bar.

24-hour boards as of 26 Oct 2026, 10:51 UTC · places follow the leaderboard’s statistical ties · the full leaderboard

• fetch: true fills the menu from GET /v1/models, which needs a valid aism_ key; the default list shows the routing modes first.
• Conversation titles use auto-fastest-quality: the fastest model among those within 5 points of the best one you can use.
• Embeddings for RAG go through POST /v1/embeddings, which requires an active OpenAI provider key.
• Text only: a request with images, audio or files gets a 400 unsupported_content.

401 invalid_api_key

Missing or invalid aism_ key. Check the key is correct and not revoked, and that the tool sends it as Authorization: Bearer aism_…. If a field insists on an sk- prefix, sk-aism_… is accepted: the prefix is stripped automatically. A key that had an expiry date and passed it gets 401 expired_api_key, and the message says when.

400 model_not_found

An unknown auto-* name is a 400 that lists the valid routing modes. For a real model id, check the spelling: direct pin routing infers the provider from the model name prefix, and GET /v1/models lists every id you can use.

No models in the dropdown

The tool fills its list from GET /v1/models, which requires a valid aism_ key. Check that the key and the base URL https://aistupidlevel.info/v1 are both in place.

400 provider_key_rejected or no_provider_keys

Your provider key is missing or was refused. Add or update it under Providers, re-check it on the provider’s dashboard, and make sure it has active credits and the right permissions.

Streaming

Set stream: true. Text arrives as the model writes it (X-AISM-Streaming-Mode: native); tool calls arrive in one chunk each once the model has finished them, then the finish reason, then usage if you set stream_options.include_usage. If the chosen model fails before sending anything, or its provider sends nothing for 120 seconds, the next fallback is tried.

400 unsupported_content

The router routes text only, on purpose: it picks a model from benchmark results for coding, reasoning and tool use, and none of those measure image, audio or file understanding. Send such requests to the provider directly.

429 allowance_exhausted

This month’s requests are used up and you have no prepaid credits. Credits top up the allowance; moving up a plan is the cheaper way to buy regular volume.