Agent tools you declare
instead of implement.

You define the schema. Charter does the request, the auth and the wire format.

pip install charter

549 tools · 15 APIs · 2 dependencies · Apache 2.0

What you write, and the
request it makes.

Four markers carry the mechanics. There is no per-endpoint code in Charter, generated or hidden.

tools/gmail.py

from typing import Annotated
from pydantic import BaseModel
from charter import Path, Query

class ListMessages(BaseModel):
    user_id: Annotated[str, Path()] = "me"
    q: Annotated[str, Query()]
    max_results: Annotated[int, Query()] = 10

list_messages = gmail(
    name="list_messages",
    method="GET",
    url_template="gmail/v1/users/{user_id}/messages",
    args_schema=ListMessages,
)

on the wire

# the model supplies only this
{"q": "is:unread", "max_results": 5}

GET /gmail/v1/users/me/messages
    ?q=is%3Aunread&maxResults=5
Host: gmail.googleapis.com
Authorization: Bearer ya29.a0Ae4lvC...

When a call fails it was the arguments or it was the API. There is no tool logic for it to be. Charter describes one request: pagination, retries and multi-call compositions belong in your agent.

Same model. Same prompt.
A different tool surface.

One triage task, temperature 0, thirty generations per cell. The only thing that changed is how much of Linear's filter the tool exposed.

the filter the model was given	bytes	glm-5p3-flash	deepseek-v4p1	nemotron-lightning
removed entirely	1,131	0/30	0/30	0/30
the full mirror	187,655	22/30	400 ×30	400 ×30
curated	15,531	28/30	30/30	4/30

The capability was never missing. The shape was.

The 400s are the provider refusing the request: the filter refers back to itself, and the cycle resolves past the depth limit before the model sees anything. The same model used the filter 21/21 times when the schema was accidentally flat, and 0/40 after.

Against the glue people write first.

534 runs on live accounts, no mocks. Two arms over the same tasks, models and prompts, differing only in the tool surface.

across 534 measured runs	raw HTTP tool	charter
malformed GraphQL documents	32	0
calls to endpoints never declared	10	0
base64 the API rejected	4	0
response bytes forwarded per run	36,383	7,382

The raw arm’s endpoint list is derived from the pack’s own tools, so nobody hand-picked what it could reach. The published record includes the tasks where the result was a wash, and the one template that goes the other way.

Full methodology →