The Crow Programming Language
SpecGitHub

A language model as a connection: use llm

reach makes a model a connection — anthropic://<model>, or openai://<host>/<model> for any server speaking the OpenAI chat completions (Ollama, vLLM, OpenRouter, LiteLLM). The key is read from the environment when a call is made; a local host needs none. What reach answers is a description the program can read:

1
2
3
4
5
6
7
8
9
10
use llm

m = reach['openai://localhost:11434/llama3.2']
log[m.provider]
log[m.model]
log[m.base]
log[m.key]
bad = reach['ftp://somewhere']
log[bad.ok]
log[bad.reason]

expected output — from the book, verified by crow test --book

openai
llama3.2
http://localhost:11434
OPENAI_API_KEY
false
cannot reach 'ftp://somewhere': a model's URL is anthropic://<model> or openai://<host>/<model>

A conversation is a parsed value, a list of turns, built a turn at a time: ask the user's, said the model's reply, told a tool's result. A reply is a parsed value too — its text, its calls, its usage and its stop — so the builders take what chat answers:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
use llm

talk = ask['What is the capital of Mongolia?']
log[len[talk]]
log[talk[0].role]
talk = ask[talk, 'And its population?']
log[len[talk]]
r = json['{"text": "Let me check.", "calls": [{"id": "c1", "name": "lookup", "args": {"city": "Ulaanbaatar"}}], "usage": {"input": 12, "output": 7}, "stop": "tool"}']
talk = said[talk, r]
log[talk[2].role]
log[talk[2].calls[0].name]
talk = told[talk, r.calls[0], '1.6 million']
log[talk[3].role]
log[talk[3].id]

expected output — from the book, verified by crow test --book

1
user
2
assistant
lookup
tool
c1

chat asks the model; a server nobody answers is the call's error value, of kind llm, and so is a stream that cannot be opened (follow), whose pieces (next) and whole (reply) then fail with it. embed answers a matrix of embeddings from an OpenAI-compatible endpoint, which Anthropic does not serve:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
use llm

m = reach['openai://127.0.0.1:1/llama3.2']
talk = ask['Hello?']
r = chat[m, talk, system = 'Answer in one word.', max = 16]
log[r.ok]
log[r.kind]
s = follow[m, talk]
log[s.ok]
piece = next[s]
log[piece.ok]
whole = reply[s]
log[whole.ok]
e = embed[m, ['one', 'two']]
log[e.ok]
a = reach['anthropic://claude-sonnet-5?key=CROW_BOOK_NO_SUCH_KEY']
log[a.reason]

expected output — from the book, verified by crow test --book

false
llm
false
false
false
false
cannot reach 'anthropic://claude-sonnet-5?key=CROW_BOOK_NO_SUCH_KEY': the key's variable CROW_BOOK_NO_SUCH_KEY is not set

Tools are the functions themselves: tools = [weather] hands the function, and the compiler writes its schema — the name, the parameters with the kinds the body reads them as, and the comment above the function as the description. The program dispatches the model's calls and hands each result back with told:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
use llm

--- the current weather in a city, in °C
fn weather[city]
    return 'sunny in `upper[city]`'

m = reach['openai://127.0.0.1:1/llama3.2']
talk = ask['How warm is it in Lisbon right now?']
r = chat[m, talk, tools = [weather]]
while r.ok and r.stop eq 'tool'
    talk = said[talk, r]
    for call in r.calls
        if call.name eq 'weather'
            talk = told[talk, call, weather[call.args.city]]
    r = chat[m, talk, tools = [weather]]
log[r.ok]
log[weather['Lisbon']]

expected output — from the book, verified by crow test --book

false
sunny in LISBON