Look up how many tokens a model can work with at once, without calling the model.
dotnet tool install -g RecursiveAI.OneLlm.CliRequires .NET 10 (net10.0); .NET 9 and earlier aren’t supported.
onellm catalog gpt-6-lunaUnzip the example, then dotnet run.
Replace gpt-6-luna with any model name from the catalog. To list the names, run onellm catalog with no arguments.
The tool prints the model’s fields, including these lines. [shipped] means the value came from the catalog embedded in the package.
model: openai:gpt-6-luna provider: openai name: "gpt-6-luna" [shipped] displayName: "GPT-6 Luna" [shipped] kind: "chat" [shipped] contextWindow: 1050000 [shipped] maxOutputTokens: 128000 [shipped]
If the name matches more than one provider, the tool lists the matches. Use provider:model to choose one.
The full program
The full program needs Example.csproj from the zip: the project sets the CatalogTool assembly attribute that the program reads to find the CLI. The download uses the same catalog command through the pinned RecursiveAI.OneLlm.Cli package. dotnet run restores the tool package and runs it without a global tool install. It prints the entry shown above, followed by capability and pricing fields. Only the package restore needs NuGet access or a populated package cache.
using System.Diagnostics; using System.Reflection; // Run the same catalog command as the installed onellm tool. var tool = Assembly.GetExecutingAssembly() .GetCustomAttributes<AssemblyMetadataAttribute>() .Single(item => item.Key == "CatalogTool").Value!; var start = new ProcessStartInfo { FileName = Environment.GetEnvironmentVariable("DOTNET_HOST_PATH") ?? "dotnet", ArgumentList = { tool, "catalog", "gpt-6-luna" } }; using var process = Process.Start(start)!; await process.WaitForExitAsync(); return process.ExitCode;
Variations
List the catalog names and windows together. Without a model name, the tool prints one tab-separated row per entry. Find the model in the first column and read its window from the second.
onellm catalogThe list includes this row:
openai:gpt-6-luna 1050000Related recipes
The short answer
Install the RecursiveAI.OneLlm.Cli tool, then run onellm catalog with a model name. Read contextWindow for the window and the bracketed label for where the value came from. The lookup reads local catalog data rather than querying a provider endpoint. It doesn’t tell you whether your account or deployment can use that window.
How does the LLM context window work?
The context window is a limit on what the model can process together, not a count of characters in the last message. When planning a request, account for the system prompt and earlier turns you send along with the new prompt and documents. The OneLLM counting docs describe allowances for messages, tools, images and other files in addition to text.
contextWindow and maxOutputTokens describe different limits. The first is the model’s catalog window; the second is a cap on generated output. Neither says how many tokens your current conversation occupies. Some entries also have a separate input limit; the example prints it as properties.maxInputTokens farther down. Don’t size a prompt using the window alone: check the input and output limits for the endpoint you actually use.
What are the context window limitations for LLMs?
The terminal command tells you what OneLLM knows locally, not how your particular server is configured. If your deployment imposes a smaller window than the catalog lists, use the deployment’s documented limit. If you maintain catalog overrides, check which layer supplied contextWindow before assuming everyone running the same command will see the same number.
The catalog specification defines three sources: shipped for the embedded data, user for user files, and code for overrides set in code. Higher layers replace individual fields, so an override of the window doesn’t have to replace the model’s other settings. The source label identifies the winning layer, not the vendor documentation or a measurement of a live run.