Ax for Python

Build Ax programs from Python without giving up the Ax model: typed signatures, structured generation, provider routing, RLM agents, flows, and optimizer artifacts all come from the same shared compiler contract. The package feels like Python, but the behavior stays aligned with the main Ax implementation.

Quick Start

pip install axllm

Realtime audio over WebSocket is an opt-in extra (pulls websocket-client):

pip install axllm[realtime]

from axllm import s

sig = s("question:string -> answer:string")
schema = sig.to_json_schema("outputs")
assert "answer" in schema["properties"]

What You Can Build

• Signatures and schemas: describe inputs and outputs once, then reuse that shape for validation, prompts, tools, and typed results.
• AxGen: run structured generation with retries, tool calls, field processors, assertions, traces, usage, and provider-backed output parsing.
• AxAI: select named deployment profiles independently from model IDs, resolve model-aware structured-output modes, and attach renewable request credentials through one provider boundary.
• Audio and realtime: .chat() accepts input_audio content parts, transcribe()/speak() do batch speech-to-text and text-to-speech, and realtime-capable models stream audio over a WebSocket — transparently through chat() or via the productized realtime_chat() driver (Go: RealtimeChat).
• Meta Muse: meta (Responses), meta-chat (Chat Completions), and meta-messages (Messages) default to muse-spark-1.3 with MODEL_API_KEY. Muse Image uses chat and results[].images; Muse Voice uses batch transcription or realtime streaming chat. Contributor models expose provider-training data-use metadata and are never defaults. Glimmer uses the existing self-hosted profiles; Ax does not manage weights.
• Jev and OpenAI Decisions: use required boolean/class/rubric score signatures with (evidence) to retain probabilities, native values, optional confidence, and trusted provenance through AxGen. Generative models use the same contract with elicited evidence. Rerank query/document pairs by the continuous rubric score while preserving candidate IDs.
• Native decision clients: use explicit provider request shapes for advanced criteria, structured state, ordered questions, and provider-specific media. Normal evidence-bearing signatures support scores and answered/refused outcomes without calling these clients directly.
• AxAgent and RLM: let an agent plan and execute actor-code steps while Ax keeps envelopes, state, logs, traces, context, discovery, recall, and final typed responses aligned.
• AxFlow: compose AxGen, AxAgent, and nested flows into a portable program graph.
• Optimizers: save, load, apply, and evaluate optimizer artifacts, including the generated GEPA engine.

Package Shape

• Import package: axllm
• Distribution metadata: pyproject.toml, MANIFEST.in, and axllm/py.typed
• Base dependencies: none
• Network support: available

Shared Ax behavior is Core-owned. The generated target code stays focused on idiomatic wrappers, transports, dynamic value helpers, and host-runtime boundaries. Authentication-required profiles accept a static key or a credential callback; the callback is invoked on every request attempt and its headers override static profile authentication.

Examples

no-key examples are deterministic local smokes. They are the fastest way to see the package work without any provider account:

• python examples/signature_schema.py: signature parsing and JSON schema generation
• python examples/model_catalog.py: model catalog, named profiles, thinking levels, and service tiers
• python examples/axgen_scripted_client_tool.py: AxGen with a scripted client, a tool, and a caching function
• python examples/provider_mapping_no_key.py: provider mapping through a scripted transport
• python examples/runtime_hooks_no_key.py: runtime-hook globals and typed program surfaces through a scripted transport
• python examples/adaptive_balancer_no_key.py: adaptive balancer state, scoring, and stable route keys without a provider key
• python examples/provider_stream_no_key.py: provider streaming through a scripted SSE transport
• python examples/cancellation_no_key.py: one-shot provider cancellation with no transport attempt
• python examples/axflow_program_graph.py: AxFlow program graph and a caching function, whose hits run no node and record no span
• python examples/flow_mermaid.py: portable Mermaid flow parsing and canonical round-trip
• python examples/audio_responses_mapping.py: OpenAI Responses speak/transcribe mapping through a scripted transport
• python examples/realtime_audio_events.py: Grok/Gemini realtime audio setup, input, and event folding
• python examples/realtime_audio_turn.py: drive a full realtime audio turn through the productized realtime_chat() driver (offline, scripted transport)
• python examples/runtime_adapter.py: custom AxCodeRuntime session
• python examples/runtime_protocol.py: process runtime protocol against the AxJS reference adapter
• python examples/optimizer_artifact.py: optimizer artifact save/load/apply lifecycle
• python examples/gepa_local_optimizer.py: local GEPA optimizer artifact generation
• python examples/ace_playbook.py: grow an evolving context playbook with playbook() (offline, scripted client)
• python examples/agent_playbook.py: attach a seeded agent playbook, exercise stage instructions and citations, learn from run-end failures, and verify accept/rollback evolution (offline, scripted client)
• python examples/mcp_scripted_tools.py: MCP tool discovery and invocation through a scripted transport
• python examples/mcp_modern_roundtrip.py: modern MCP discovery, cache, task, and roots MRTR over an in-process HTTP loopback
• python examples/context_cache_recovery.py: Gemini managed-context-cache create, refresh/recreate, rejection invalidation, and uncached fallback

provider-api examples make a real provider call. OpenAI examples require OPENAI_API_KEY; each Vertex command below lists its native-routing or OpenAI-compatible endpoint variables:

• OPENAI_API_KEY=... python examples/axgen_openai_api.py: GPT-5.6 prompt-cached AxGen with the OpenAI Chat API
• GOOGLE_VERTEX_ACCESS_TOKEN=... GOOGLE_PROJECT_ID=... GOOGLE_REGION=... python examples/vertex_gemini_api.py: Gemini through Vertex routing
• OPENAI_API_KEY=... python examples/flow_openai_api.py: AxFlow with a real OpenAI-compatible provider API

Runtime Profiles And RLM Agents

AxAgent uses an RLM executor loop. On each turn, the model writes a small actor-code step, and Ax sends that step into an AxCodeRuntime session. Think of the runtime as the agent's REPL: it keeps session state, exposes safe host callbacks, returns envelopes such as final(...), askClarification(...), discover(...), recall(...), and used(...), and lets the agent continue from the result.

The TypeScript package ships AxJSRuntime as the reference JavaScript implementation of that REPL contract. Generated runtime profiles are adapters for the same AxCodeRuntime / AxCodeSession boundary. They exist so RLM agents can execute actor code in a host runtime that fits the target package.

This package is not a TypeScript transpiler. AxIR compiles shared Ax semantics into native package code; it does not run your original Ax TypeScript application inside a Python runtime. Application code is still written in the language you are using here.

Optional profile files in this package:

• javascript-quickjs: JavaScript actor code through a QuickJS protocol server via ProcessCodeRuntime.
• python-pyodide: Python actor code through a Pyodide JSONL protocol server.

See examples/runtime_profiles/README.md for setup, policy, and verification details.

Optional runtime profiles are dependency-bearing and opt-in. Adapter policy owns sandboxing, dependency loading, hard cancellation, process security, and host permissions. The shared Ax contract still owns envelopes, state, logs, traces, and the model-visible protocol.

Contract Snapshot

• Compiler contract version: 0.1
• Package: axllm
• Supported conformance suites: signature, schema, validation, prompt, axgen, axai, axagent, axoptimize, axprogram, axflow, axmcp, axevent
• Provider mode: provider-descriptor-registry-openai-compatible-openai-responses-google-gemini-anthropic
• Scripted transport support: true
• Real network support: available