Cadenza.Agent

Cadenza.Agent

Read this in 한국어.

Cadenza.Agent is the AI-agent variant of the Cadenza SDK family — a single-file scripting MSBuild SDK that wraps Microsoft.Extensions.AI (the vendor-neutral .NET LLM abstraction) and exposes an OpenAI-compatible Chat Completion HTTP endpoint out of the box. That means tools like Codex, Aider, Continue, and Cursor can talk to your agent as if it were OpenAI — just point them at http://localhost:8080/v1.

Quick start

Create an agent.cs file:

#!/usr/bin/env dotnet run
#:sdk [email protected]

SystemPrompt("You are a helpful assistant with filesystem access.");

Tool("read_file", "Read a UTF-8 text file from disk",
    (string path) => ReadText(path));

Tool("list_files", "List files matching a glob pattern",
    (string pattern) => Glob(pattern).ToArray());

UseOllama("llama3.2");      // or UseOpenAi("gpt-4o-mini")
                            //    UseAnthropic("claude-3-5-sonnet-latest")
                            //    UseAzureOpenAi(endpoint, deployment)

await Run();                // boots the OpenAI-compatible HTTP server
                            // on http://localhost:8080

Run iteratively:

dotnet run agent.cs

Point any OpenAI-compatible client at it:

export OPENAI_BASE_URL=http://localhost:8080/v1
export OPENAI_API_KEY=any-non-empty-string
codex      # or aider, continue, cursor, sgpt, …

Or use the REPL instead of the HTTP server:

UseOllama("llama3.2");
await ChatLoop();           // interactive console — no HTTP server

Publish as a self-contained single binary:

dotnet publish agent.cs -r linux-x64 -c Release

What you get

Endpoints exposed by Run()

• POST /v1/chat/completions — full OpenAI Chat Completion shape; supports stream=true (SSE). Used by Aider, Continue, Cursor, GitHub Copilot BYOK, sgpt, and almost every other OpenAI-compatible client.
• POST /v1/responses — OpenAI Responses API (SSE streaming). Used by OpenAI Codex CLI, which removed Chat Completion support in Feb 2026 and now requires this wire format exclusively.
• GET /v1/models — single-entry list with ServedModelName as the id.
• GET /health — liveness probe.

Chat Completion vs Responses — what's different

Both endpoints are backed by the same IChatClient you configured with UseOllama / UseOpenAi / etc. They differ in how tools are handled:

So if you register Tool("read_file", ...), it works for Aider/Continue/Cursor but not for Codex. For Codex, the model adapter pattern is enough — Codex brings shell / apply_patch / update_plan and executes them itself.

Connecting Codex CLI

Drop this into ~/.codex/config.toml (or %USERPROFILE%\.codex\config.toml on Windows):

model_provider = "cadenza"
model          = "cadenza-agent"

[model_providers.cadenza]
name     = "Cadenza.Agent local"
base_url = "http://localhost:8080/v1"
wire_api = "responses"
env_key  = "CADENZA_API_KEY"
stream_idle_timeout_ms = 300000

Then set any non-empty value for CADENZA_API_KEY and run Codex.

Why an HTTP frontend?

Most existing coding agents and chat UIs already speak an OpenAI wire format. Adopting both Chat Completion and Responses instead of a bespoke protocol means:

• Zero glue code to integrate with Codex / Aider / Continue / Cursor / sgpt / any LangChain client.
• Free streaming via SSE — same shape OpenAI uses.
• Multi-process by default — your agent process is decoupled from the editor.
• Function-calling is auto-wired through Microsoft.Extensions.AI's UseFunctionInvocation middleware, so any tool you register is callable by every supported model.

See the project repository for the full specification, sample agents, and security notes.