Skip to main content
Pre-release

This release is a pre-release and may not be stable for production use.

JAZ

A Python framework for building and optimizing LLM-agents through intelligent, iterative execution.

Overview

JAZ provides a unified framework for creating agents that combine deterministic Python code with LLM reasoning. Agents execute in a REPL loop where the LLM generates code, observes results, and iterates until the task is complete.

Prerequisites

  • Python ≥ 3.12 — 3.14+ is recommended for the full feature set. The t-string prompt syntax (invoke(task=t"Analyze the {data}")) and the interactive jaz console need PEP 750 t-strings, hence Python ≥ 3.14. Everything else works on 3.12/3.13; the equivalent invoke(task="Analyze the `data`", data=data) works on every supported version.

Installation

pip install jaz-lang

The distribution is named jaz-lang; the import name is jaz:

import jaz

Optional extras:

pip install "jaz-lang[tracing]"   # OpenTelemetry tracing
pip install "jaz-lang[copilot]"   # GitHub Copilot backend — no API key, billed to AI credits

To pin a specific release, or to install one that has not reached PyPI, use the distribution repo and a release tag:

pip install "jaz-lang @ git+https://github.com/jaz-lang/jaz.git@v0.2.0b1"

API Keys

JAZ reads provider API keys from the environment, under each provider's standard variable name. Set them via any of:

  • Shell profile — add export OPENAI_API_KEY=sk-... to ~/.zshrc or ~/.bashrc (simplest, available globally)
  • .env file — create a .env at the project root with OPENAI_API_KEY=sk-... per line, then load it with:
    • direnv — add dotenv to an .envrc file to auto-load on cd (brew install direnv)

No API key at hand? The copilot backend (the copilot extra) bills a Copilot subscription's GitHub AI credits instead. Authentication checks an explicit github_token first, then COPILOT_GITHUB_TOKEN / GH_TOKEN / GITHUB_TOKEN in the environment (in that order), then your signed-in gh/copilot CLI. An exported token therefore takes precedence over your CLI login:

import jaz
from jaz.llm import GitHubCopilotLLM

jaz.configure(llm=GitHubCopilotLLM(model="gpt-5-mini"))   # ids are Copilot's own, unprefixed

For CLI authentication, run gh auth login, or start copilot and enter /login.

Like every backend other than LiteLLM it sits outside the stable API surface — see Custom & Local LLM Backends.

Quick Start

from jaz import invoke
from jaz.hooks import ReturnType

# Simple task
result = invoke(
    ReturnType(int),
    task="Calculate the factorial of 10",
)

# With custom inputs
result = invoke(
    ReturnType(list),
    task="Double each number in the list",
    numbers=[1, 2, 3, 4, 5]
)

Later snippets assume the Quick Start imports (from jaz import invoke, from jaz.hooks import ReturnType, import jaz) and show only what each adds.

Every input is a keyword argument — the prompt included. invoke()'s positional arguments are hooks (ReturnType(int), a budget, …) and an optional ConfigOverride, so invoke("Calculate the factorial of 10") is an error, not a shorthand.

The jaz console

Installing puts a jaz command on PATH (it needs Python ≥ 3.14 to run): a real Python REPL — variables persist, _ holds the last result — with conversational sugar layered on top. Run jaz, then enter these lines at its prompt:

> what is 2 + 2?
n: int <- how many days in a leap year?
?n
% switch to gpt-5-mini
import math

> asks the agent; n: int <- captures a typed result; ?n shows what the agent sees for n; % asks the jaz helper to change settings, with confirmation. Plain Python still runs. Enter the sigil lines without trailing # annotations: they become prompt text, or can leave a ? expression waiting for more input.

A session with no model configured opens a setup wizard first: pick a provider and paste its API key, or pick github copilot — no provider key, just the copilot extra and GitHub authentication. The choice is validated with one live request before anything is kept. Backend and model defaults are saved to ~/.jaz/settings.json; provider API keys entered in the wizard go in ~/.jaz/credentials.json.

Flags configure one session without touching that default — jaz --backend copilot --model gpt-5-mini, jaz --max-cost 0.50, jaz --log run.log — and jaz -c "…" runs a single prompt and exits. In-session, setup_llm() re-runs the wizard and switch_model("openai/gpt-5") changes model (keeping the backend) after validating it.

Key Features

Recursive Agent Composition

Agents can invoke nested agents with separate budgets and automatic cost tracking:

from jaz.hooks import RecursionLimit, ReturnType

with RecursionLimit(max_depth=3):  # optional cap; recursion is unbounded by default
    result = invoke(
        ReturnType(list),
        task="""
        For each item, use invoke to process it.
        Combine and return the results.
        """,
        items=data,
    )

Custom Tools

A tool is just a function you pass in. Give it as a keyword argument and it binds to that name in the agent's REPL; its signature and docstring render in the prompt, so the docstring is the description the agent reads:

def my_tool(x: int) -> int:
    """Double a number."""
    return x * 2

result = invoke(ReturnType(int), task="Use my_tool on 5", my_tool=my_tool)

The agent sees `my_tool(x: int) -> int`: Double a number. — type hints and a one-line docstring are what make a tool usable. Nothing is registered up front, so tools sit alongside data in the same call:

def web_search(query: str) -> list[str]:
    """Search the web and return result snippets."""
    ...

result = invoke(
    ReturnType(dict),
    task="Look up each name and summarize what you find",
    names=["ada", "grace"],
    web_search=web_search,
)

To group related tools under one name, pass an object instead: it renders as its class docstring with one bullet per public method, and the agent calls them as utils.my_tool(...).

class Utils:
    """Utility helpers."""

    def my_tool(self, x: int) -> int:
        """Double a number."""
        return x * 2

result = invoke(ReturnType(int), task="Use utils.my_tool on 5", utils=Utils())

The agent sees:

`utils` [Utils]: Utility helpers.
  - `my_tool(x: int) -> int`: Double a number.

Data and other values render as their str(), so define __str__ on a type to control what the agent sees for it, or __repr__ for a value that sits inside a list or dict.

To make a tool propagate automatically to nested invoke calls, bind it with jaz.scope instead of passing it per call:

from jaz import scope

with scope(my_tool=my_tool):
    invoke(ReturnType(int), task="Use my_tool on 5")  # nested invokes inherit `my_tool`

Hooks for Extensibility

Use hooks to add logging, tracing, workflow strategies, and more:

from jaz.hooks import BudgetPool, FileLogger

with FileLogger("agent.log"), BudgetPool(cost_budget=1.0):
    result = invoke(ReturnType(str), task="Build a web scraper")

Budget Control

Pool budgets (LLM cost / call count, shared across the whole invoke/recursion tree) are enforced by the opt-in BudgetPool; per-level execution limits are opt-in hooks as well:

from jaz.hooks import BudgetPool, IterationLimit, RecursionLimit

# Pool budgets: tracked + enforced only while the hook is active.
with BudgetPool(cost_budget=0.50, calls_budget=100):  # USD / call-count
    result = invoke(ReturnType(str), task="Complex task")

# Per-level execution limits are hooks too:
with IterationLimit(max_iterations=20), RecursionLimit(max_depth=2):
    result = invoke(ReturnType(str), task="Complex task")

Configuration

Override defaults globally or per-invoke:

Each group takes the configured component, whose constructor is that group's settings. Setting a group replaces it — a component states itself completely, so there is no partial update of one setting.

from jaz.llm import LiteLLM
from jaz.repl import PythonREPL

# Per-invoke override
with jaz.ConfigOverride(llm=LiteLLM(model="openai/gpt-5-mini", temperature=0.7, max_tokens=2000)):
    result = invoke(ReturnType(str), task="Creative task")

# Global configuration
jaz.configure(llm=LiteLLM(model="openai/gpt-5-mini"), repl=PythonREPL(exec_timeout=60))

Custom & Local LLM Backends

JAZ talks to LLMs through a small BaseLLM layer. LiteLLM is the default — one backend that routes to every provider LiteLLM supports (OpenAI, Anthropic, Gemini, Bedrock, Vertex, …) — and GitHubCopilotLLM and SGLangLLM cover the two cases it does not serve well: Copilot billing through GitHub AI credits, and a self-hosted SGLang server addressed token-in/token-out. All three are part of the stable API. You can add your own backend without forking.

A config selects one backend: the BaseLLM you pass to llm=, defaulting to LiteLLM. Its constructor takes everything it needs — the model id (a LiteLLM route like openai/gpt-5-mini), the backend's own settings, and any per-request params. A config file can name a backend instead of constructing one, but the name lives in whatever reads the file: JAZ's own loaders (the eval harness, the console) each keep their own tag → class mapping, and core carries neither the mapping nor a name on the class. Because litellm is the default, a config need only name the model.

Local OpenAI-compatible server (Ollama, vLLM, LM Studio, llama.cpp, …) — no code, just config:

# route the litellm backend to a local OpenAI-compatible server
# (api_key is required by the openai/ route even for a local server — a dummy value is fine;
#  it rides the query params into litellm.completion)
jaz.configure(
    llm=LiteLLM(model="openai/llama3", api_base="http://localhost:11434/v1", api_key="dummy"),
)

A custom backend — subclass BaseLLM and pass an instance (a name is only needed if a config file must select it, and it is that file's reader that holds the name):

import os
from functools import cached_property
from jaz.llm import BaseLLM, LLMResponse, ModelInfo, RetryPolicy

class MyLLM(BaseLLM):
    def __init__(self, model: str, base_url: str | None = None, retry_policy=RetryPolicy(), **query_params):
        super().__init__(retry_policy=retry_policy)   # the one thing the base holds
        self._model = model                          # yours: the base has no `model`
        self._base_url = base_url or os.environ["MYBACKEND_API_BASE"]
        self._query_params = query_params            # yours too; you put them on the call

    def complete(self, messages):
        # ... call self._base_url; raise jaz.exceptions.RateLimitError / AuthenticationError / ...
        # on API errors so retry classification works.
        wire = my_http_post(self._base_url, model=self._model, messages=messages, **self._query_params)
        return LLMResponse(
            content=wire["text"],
            prompt_tokens=wire["usage"]["input"],
            completion_tokens=wire["usage"]["output"],
            cost_usd=None,                    # or price it yourself; see below
            extra={},
        )

    @cached_property
    def model_info(self):
        return ModelInfo(name=self._model)    # what you know; every field defaults to "unknown"

jaz.configure(llm=MyLLM(model="my-model", base_url="http://localhost:8000"))

One class owns the whole job: the API call, its parameters, and whatever accounting you choose to report. complete() receives the messages and returns an LLMResponse — that is the entire contract, and BaseLLM, LLMResponse, ModelInfo and RetryPolicy are the only names you need from jaz.llm. Nothing passes through the base: the model, the sampling parameters and the credentials are your own state, set in your __init__ and put on the request by your complete. The base holds retry alone, so a backend with a **query_params tail names retry_policy explicitly and forwards only that. The base has no model: a backend that serves a nameable model takes model= as its own setting (all three built-ins do) and reports it through model_info.name, which is what tracing hooks label spans with; one that serves whatever its server loaded leaves the name None. The retry wrappers come from the base, which retries a failed call only on its retryable_exceptions (RateLimitError, ServerError and APIConnectionError, timeouts included, from jaz.exceptions by default): raise a transient failure as one of those, or add your own class to that tuple, and anything else fails at once. A user can retry more errors with RetryPolicy(extra_retryable_exceptions=...).

Cost is yours to report. Set cost_usd if you can price the call — LiteLLM's pricing helpers (litellm.completion_cost, litellm.model_cost) are what the built-in LiteLLM backend prices with, and litellm is a JAZ dependency, so it is already installed.

If your backend can price calls, say so with ModelInfo(can_report_cost=True) — the default is False, because JAZ itself prices nothing. Leaving it alone is safe: a cost_budget is then rejected before any call with an error naming your backend, and calls_budget still works. max_input_tokens is the same shape — it defaults to None ("unknown", not "unlimited"), and the ContextWindowWarning hook sizes its window from it, so a backend that does not report one just leaves that hook disabled rather than erroring. Retry is one value too: MyLLM(model=..., retry_policy=RetryPolicy(max_retries=2)).

When a config file names your backend, every key in its block is a constructor argument: model and base_url reach the named parameters, and anything else (temperature) lands in your **query_params tail — a backend declares its settings just by declaring __init__. Whoever loads that file supplies the name → class mapping (the eval harness does it with the non-public jaz._instantiate.build_config(..., resolvers={"llm": my_lookup})); JAZ core holds no registry, and a backend class carries no name of its own. Constructing it yourself needs no name at all: ConfigOverride(llm=my_llm).

Testing without a provider. MockLLM is a backend that answers from a callable, so an agent can be exercised with no API key, no network and no nondeterminism — through the same configure/ConfigOverride as any backend, with retry and cost accounting intact:

from jaz.llm import MockLLM

def scripted(messages):
    return "return 42"                    # a str is the assistant's reply; or return an LLMResponse

with jaz.ConfigOverride(llm=MockLLM(scripted)):
    assert jaz.invoke(task="add 40 and 2") == 42

LiteLLM is the default; GitHubCopilotLLM and SGLangLLM ship alongside it for the cases it does not cover, and all three are covered by the stable API. Reach OpenAI and Anthropic through litellm's openai/… / anthropic/… routes — the hand-written vendor-direct backends were removed before the first release. To add or replace a backend, subclass BaseLLM; add it to your loader's map only if a config file needs to name it.

The configured model id is sent verbatim: JAZ neither strips nor rewrites a provider/ prefix, so write the id the backend's own documentation uses. A backend is chosen by configuring it, not by prefixing the model — the prefix tells the configured backend which provider to route to, and a backend that does not route (SGLangLLM) sends the id to its own server verbatim.

Architecture

jaz/
├── _invoke.py      # invoke() / ainvoke(), re-exported from `jaz`
├── _config.py      # configure() / ConfigOverride / get_config()
├── _agent.py       # The REPL loop that drives one invoke
├── repl/           # REPL implementations (PythonREPL)
├── hooks/          # Hook system: events, effects, dispatcher, built-in hooks (BudgetPool, ...)
├── llm/            # LLM backends (LiteLLM by default)
└── protocol/       # Wire-format codec between the LLM and the REPL

Modules with a leading underscore are non-public; everything supported is exported from jaz or one of its public modules (jaz.hooks, jaz.llm, jaz.repl, jaz.protocol, jaz.exceptions, jaz.credentials).

Core Concepts

  • Agent loop: Orchestrates LLM queries and REPL execution
  • REPL: Executes agent-generated code with safety controls
  • Hooks: Event-based system for logging, budget control, workflow capture, and extensibility
  • Tools: Plain functions passed as inputs; their signature and docstring become the description the agent reads

Built-in Hooks

Hook Purpose
ReturnType Declare + enforce the invoke's return type
ValidateReturn / ValidateREPLCode Validate the return value / veto REPL code before it runs
BudgetPool Shared LLM cost / call-count budget with hard-stop enforcement
IterationLimit / RecursionLimit Per-level turn cap / invoke-nesting cap
BudgetForcing Refuse early finishes so the agent keeps working
ContextWindowWarning Warn the agent as its prompt nears the model's window
PrintLogger / FileLogger Log events to console / file
TrajectoryRecorder Write the run as an ATIF trajectory (replay/cost source)
TrajectoryDirectoryRecorder Write the run as a browsable Markdown directory, live (tail -f a running agent)
TrajectoryReplay Resume a run from an ATIF trace: saved responses replay, then live calls take over
RolloutRecorder Record token-native rollouts for training
JaegerTracing / LangfuseTracing OpenTelemetry tracing presets

Writing Custom Hooks

Hooks observe agent execution via events (from jaz.hooks.events) and influence it by returning effects (from jaz.hooks.effects):

from jaz import invoke
from jaz.hooks import Hook
from jaz.hooks.effects import AddMessages, Effect
from jaz.hooks.events import LLMQueryEnter

class ConciseHook(Hook):
    """Instruct the agent to be concise after iteration 3."""

    # Override the typed per-event handler for the event you care about — no
    # isinstance/dispatch boilerplate. (A cross-cutting observer that wants *every*
    # event overrides `on_any` instead.)
    def on_llm_query_enter(self, event: LLMQueryEnter) -> list[Effect]:
        if event.iteration >= 3:  # iterations are 0-based
            return [AddMessages([{"role": "user", "content": "Be concise."}])]
        return []

# Hooks activate via context managers and propagate to nested invoke() calls
with ConciseHook():
    result = invoke(task="Solve this step by step")

Hooks compose naturally as context managers:

# Stack multiple hooks
with ConciseHook(), FileLogger("agent.log"):
    result = invoke(...)

Event Types

Events fire around three spans — the whole Invoke, each turn's LLMQuery, and each turn's REPLExec — and every span walks the same four stages. All are importable from jaz.hooks.events.

Stage Observes Accepts Fires
*Enter the proposal edit effects + Abort always
*Send the committed input supply effects + Abort iff the input committed
*Complete the raw result modify effects + Abort iff a raw result exists
*Exit the outcome union (Completed | Aborted | Failed) nothing (observation-only) whenever the span opened

Plus LLMQueryRetry, fired per retry attempt of an LLM call. Every event carries contextual data (e.g. LLMQueryEnter.iteration, event.config) and a timestamp (its emission time — durations are arithmetic over two events' timestamps).

Effect Types

Effects are returned by hooks to influence execution. All are importable from jaz.hooks.effects; each composes at a specific stage (an effect returned anywhere else raises InvalidEffectError).

Effect Stage Influence
AddInputs / DropInputs InvokeEnter Add/un-pass invoke inputs (prompt + REPL)
DisableRecursion InvokeEnter Withhold the recursive invoke tool
SupplyInvokeResult InvokeSend Supply the invoke's result, skipping the agent loop
ModifyInvokeResult InvokeComplete Replace the invoke's result
AddMessages / DropMessages LLMQueryEnter Edit the messages sent to the model (transient or persistent)
SupplyLLMResponse LLMQuerySend Supply the response, skipping the API call
ModifyLLMResponse LLMQueryComplete Replace the model's response
AddToNamespace / DropFromNamespace REPLExecEnter Bind/unbind REPL namespace names for the turn
InsertCode / DeleteCode REPLExecEnter Edit the code about to run
SupplyExecResult REPLExecSend Supply a result, skipping execution
ModifyExecResult REPLExecComplete Replace the raw execution result
Abort any Enter/Send/Complete, and LLMQueryExit / REPLExecExit Abort the invoke; its error propagates out of invoke()
BlackboardWrite any event Write to the per-invoke cross-hook blackboard

Langfuse Tracing Quick Start

from jaz import invoke
from jaz.hooks import LangfuseTracing, ReturnType

with LangfuseTracing():
    result = invoke(ReturnType(int), task="Calculate 2+2")

Environment variables:

  • LANGFUSE_PUBLIC_KEY
  • LANGFUSE_SECRET_KEY
  • LANGFUSE_HOST (optional, default: https://cloud.langfuse.com)

Contributing

This repository is generated: each release publishes a cleaned copy of the package, so changes are not made here. Contributor setup and code conventions live with the source.

License

See LICENSE file.

Metadata

Release files for jaz-lang 0.2.0b1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for jaz-lang 0.2.0b1
File Size Uploaded Provenance Actions
jaz_lang-0.2.0b1.tar.gz 744.7 kB With provenance Details for jaz_lang-0.2.0b1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for jaz-lang 0.2.0b1
File Interpreter ABI Platform Actions
jaz_lang-0.2.0b1-py3-none-any.whl Python 3 none any Details for jaz_lang-0.2.0b1-py3-none-any.whl

Total release size: 1.5 MB

Release files / jaz_lang-0.2.0b1.tar.gz

Download URL jaz_lang-0.2.0b1.tar.gz
Size 744.7 kB
Tags Source
SHA-256 checksum
How to use checksums
e3b81093635bea97805e4612568de08a7d1bbce1c77f98208745a94f2d87fcfc
BLAKE2b-256 checksum
How to use checksums
4706213eef6d189b6c13b27567b45a33cc9b395e5b4c1d14127c3085c5d57d98
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 22, 2026.

Transparency log

Release files / jaz_lang-0.2.0b1-py3-none-any.whl

Download URL jaz_lang-0.2.0b1-py3-none-any.whl
Size 794.2 kB
Tags Python 3
SHA-256 checksum
How to use checksums
0f50da5fdedd836e4a62939164db3a0ba6892340013fef16a031ee993b751442
BLAKE2b-256 checksum
How to use checksums
79437f7d77794976d467037b823e1319cd7e8e1d8f83cbbbe62521452813e4ec
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 22, 2026.

Transparency log

Security for jaz-lang 0.2.0b1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for jaz_lang-0.2.0b1.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 22, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page