Serve an HTTP "web app" that lives entirely inside a Cline SDK agent's conversation history. The agent is given a prompt describing the app (e.g. "a TODO list with HTML pages and a JSON API at /api/todos"), then deepserve opens a TCP port and pipes every raw HTTP request into agent.continue(). The agent replies with a raw HTTP response, which is written straight back to the socket.

This is a toy / demo — there is no context pruning, so long sessions will eventually overflow the model's context window.

Requires Node.js 22+ or Bun.

bun install
export CLINE_API_KEY="cline_..."

# Inline prompt
bun run dev --port 8080 -- "A retro guestbook: GET / shows a form and \
the latest 10 entries; POST / accepts name and message form fields and \
appends an entry; entries are stored in your conversation memory."

# Or load the prompt from a file
bun run dev --port 8080 --prompt-file ./prompts/guestbook.txt

Then in another shell:

curl -v http://127.0.0.1:8080/
curl -v -d 'name=Ada&message=Hello' http://127.0.0.1:8080/

Each curl is one turn in the agent's conversation.

-p, --port <n>          TCP port (default 8080)
    --host <addr>       Bind address (default 127.0.0.1)
-f, --prompt-file <p>   Read the initial prompt from a file
    --provider <id>     SDK provider id (default "cline")
    --model <id>        Model id (default "anthropic/claude-sonnet-4.6")

Anything after -- (or any leftover positional args) becomes the initial prompt.

1. Construct an Agent from @cline/sdk with a minimal system prompt that says: first turn → acknowledge the spec in prose; every later turn → respond with raw HTTP/1.1 only.
2. agent.run(initialPrompt) — the model "boots" the app.
3. A net.createServer accepts one HTTP request per connection (Content-Length framed), feeds the raw bytes to agent.continue(), and writes the agent's outputText back to the socket.
4. Requests are serialized so the conversation history stays linear.