Overview

Relevant source files

• CHANGELOG.md
• README.md
• package-lock.json
• package.json
• src/aisdk/package.json
• src/aws-bedrock/package.json
• src/ax/dsp/program.ts
• src/ax/index.ts
• src/ax/package.json
• src/ax/prompts/agent.ts
• src/examples/function.ts
• src/examples/package.json
• src/tools/package.json

Ax is a TypeScript framework for building reliable AI applications with Large Language Models (LLMs). Inspired by Stanford's DSPy (Declarative Self-improving Language Programs), Ax separates what you want (typed input/output contracts called signatures) from how to achieve it (automatic prompt generation and optimization). Instead of manually engineering prompts, developers define signatures like "question:string -> answer:string" using the AxSignature class src/dsp/sig.ts300-364 and the framework handles prompt construction via AxPromptTemplate src/dsp/prompt.ts26-207 response parsing through extractValues() src/dsp/generate.ts1005-1193 and automatic performance optimization.

Core Philosophy: Treat LLM interactions as programmable modules with strong contracts, not fragile text prompts. The framework provides compile-time type safety, automatic validation, and self-improvement capabilities that eliminate the need for manual prompt iteration.

Sources: README.md1-11 src/dsp/sig.ts300-364 src/dsp/prompt.ts26-207 src/dsp/generate.ts1005-1193

What Makes Ax Different

Traditional LLM development requires manual prompt engineering, fragile text parsing, and provider-specific code. Ax eliminates these pain points through:

1. Declarative Signatures: Define typed contracts using the AxSignature system src/dsp/sig.ts300-364 - specify inputs and outputs, not prompts
2. Automatic Execution: Framework handles prompt generation via AxPromptTemplate src/dsp/prompt.ts26-207 response parsing through extractValues() src/dsp/generate.ts1005-1193 and validation src/dsp/sig.ts828-891
3. Provider Abstraction: Single AxAIService interface src/ai/types.ts works across 15+ LLM providers with automatic feature detection
4. Self-Optimization: Programs improve automatically through optimizers: AxMiPRO src/dsp/optimizers/miproV2.ts AxGEPA src/dsp/optimizers/gepa.ts and AxACE src/dsp/optimizers/ace.ts
5. Production-Ready Infrastructure: Built-in OpenTelemetry tracing, metrics collection, rate limiting, and context caching

Sources: src/dsp/sig.ts300-891 src/dsp/prompt.ts26-207 src/dsp/generate.ts1005-1193 src/ai/types.ts src/dsp/optimizers/

Core Programming Model

From Signature to Execution

The framework provides five core abstractions:

Abstraction	Class	Purpose	Factory Function
Signature	AxSignature	Type-safe input/output contract	ax(), f() src/dsp/template.ts18-62 src/dsp/sig.ts1097-1157
Program	AxGen	Executable unit implementing signature	ax() returns AxGen src/dsp/generate.ts77-240
Agent	AxAgent	Program with function calling (ReAct)	agent() src/prompts/agent.ts682-731
Workflow	AxFlow	Multi-step orchestration with state	flow() src/flow/flow.ts40-580
Service	AxAI	Unified interface to LLM providers	ai() src/ai/wrap.ts300-346

Sources: README.md1-236 src/dsp/sig.ts300-364 src/dsp/sig.ts1097-1157 src/dsp/generate.ts77-240 src/ai/types.ts src/dsp/template.ts18-62 src/prompts/agent.ts682-731 src/flow/flow.ts40-580 src/ai/wrap.ts300-346

Layered Architecture

Ax is organized into seven conceptual layers, from user-facing abstractions down to infrastructure:

Framework Architecture Layers

Layer Responsibilities:

Layer	Components	Responsibilities
User Applications	Application code, examples	Entry point for building AI applications
High-Level Abstractions	AxAgent, AxFlow, AxProgram, AxGen	Programmable units with typed interfaces
DSP Layer	AxSignature, AxPromptTemplate, extractors, assertions	Signature-driven prompt engineering and validation
AI Service Abstraction	AxAI, AxBaseAI, routers	Unified interface to LLM providers
Provider Implementations	15+ provider implementations	Provider-specific API transformations
Optimization Layer	AxMiPRO, BootstrapFewShot, AxACE, AxGEPA	Automatic program improvement
Infrastructure	API calls, memory, response processing, errors	Core utilities and state management
Supporting Systems	Telemetry, logging, caching	Observability and performance optimization

Automatic Operations: The framework handles these operations without developer intervention:

Operation	Implementation	Purpose
Prompt Generation	AxPromptTemplate.render() src/dsp/prompt.ts109-207	Converts signatures to provider-specific prompts
Response Parsing	extractValues() src/dsp/generate.ts1005-1193	Extracts typed fields from LLM responses
Type Validation	AxSignature.validate() src/dsp/sig.ts828-891	Ensures inputs/outputs match signature types
Streaming	streamingExtractValues() src/dsp/generate.ts699-926	Incremental parsing during streaming responses
Error Correction	assertAssertions() src/dsp/generate.ts928-1003	Automatic retry with fixing instructions
Optimization	AxBaseOptimizer.compile() src/dsp/optimizer.ts651-723	Automatic program tuning with examples

Sources: src/prompts/agent.ts178-513 src/flow/flow.ts40-580 src/dsp/program.ts18-227 src/dsp/generate.ts77-1193 src/dsp/sig.ts300-891 src/dsp/prompt.ts26-207 src/dsp/asserts.ts src/ai/wrap.ts69-181 src/ai/base.ts148-421 src/ai/router.ts src/ai/balance.ts src/ai/openai/api.ts src/ai/anthropic/api.ts src/ai/google-gemini/api.ts src/dsp/optimizers/miproV2.ts src/dsp/optimizers/bootstrapFewshot.ts src/dsp/optimizers/ace.ts src/dsp/optimizers/gepa.ts src/util/apicall.ts src/mem/memory.ts src/dsp/processResponse.ts src/dsp/loggers.ts

Package Structure

The monorepo contains six npm packages:

Monorepo Packages

Package	Version	Location	Dependencies
@ax-llm/ax	16.0.9	src/ax/	@opentelemetry/api@^1.9.0, dayjs@^1.11.13
@ax-llm/ax-tools	16.0.9	src/tools/	@ax-llm/ax
@ax-llm/ax-ai-sdk-provider	16.0.9	src/aisdk/	@ax-llm/ax, ai@^5.0.100, @ai-sdk/provider-utils@^3.0.10
@ax-llm/ax-ai-aws-bedrock	16.0.9	src/aws-bedrock/	@ax-llm/ax, @aws-sdk/client-bedrock-runtime@^3.709.0
@ax-llm/ax-examples	11.0.61	src/examples/	@ax-llm/ax, @ax-llm/ax-tools, OpenTelemetry SDK
@ax-llm/ax-docs	-	src/docs/	Astro, React, Tailwind

The workspace uses npm workspaces package.json71-74 with unified tooling: TypeScript 5.8+ package.json61 Vitest 3.2+ package.json62 for testing, Biome 2.1+ package.json40 for linting/formatting, tsup 8.5+ package.json56 for bundling. Minimum Node.js version is 20 package.json37-38

Sources: package.json1-78 src/ax/package.json1-70 src/aisdk/package.json1-47 src/tools/package.json1-46 src/examples/package.json1-41 src/aws-bedrock/package.json1-54

Request Execution Pipeline

When you call program.forward(ai, input), the request flows through six stages:

Core Execution Flow - From User Code to LLM Response

Pipeline Stages:

1. Input Validation src/dsp/sig.ts828-891: AxSignature.validate() checks types and required fields
2. Prompt Generation src/dsp/prompt.ts109-207: AxPromptTemplate.render() formats signature, examples, and demos with cache flags
3. Memory Update src/mem/memory.ts63-270: AxMemory.addRequest() stores user message in conversation history
4. AI Service src/ai/base.ts220-421: AxBaseAI.chat() applies rate limiting, tracing, and delegates to provider
5. Provider Transformation src/ai/openai/api.ts192-511: Provider-specific createChatReq() and createChatResp() methods
6. Response Extraction:
   • Streaming: streamingExtractValues() src/dsp/generate.ts699-926 with incremental parsing
   • Non-streaming: extractValues() src/dsp/generate.ts1005-1193 with complete parsing
7. Validation & Assertions src/dsp/generate.ts928-1003: assertAssertions() runs checks, retries on failure
8. Function Calling src/dsp/functions.ts205-457: AxFunctionProcessor executes tools and loops if needed

Automatic Retry Logic: If assertions fail at stage 7, the framework automatically:

• Adds error correction instructions to memory
• Retries the request (up to maxRetries times, default 3)
• Maintains conversation context across retries

Sources: src/dsp/generate.ts196-1193 src/dsp/sig.ts828-891 src/dsp/prompt.ts26-207 src/mem/memory.ts63-270 src/ai/base.ts220-421 src/ai/openai/api.ts192-511 src/dsp/functions.ts205-457 src/ai/types.ts

Provider Ecosystem

Ax abstracts 15+ LLM providers behind a unified AxAIService interface src/ai/types.ts:

Multi-Provider Architecture

Provider Implementation Pattern: Each provider extends AxBaseAI and implements:

• createChatReq() - Transform AxChatRequest to provider format
• createChatResp() - Parse provider response to AxChatResponse
• createChatStreamResp() - Handle streaming responses

Advanced Routing Options:

• AxBalancer src/ai/balance.ts40-299: Load balancing with automatic failover and exponential backoff
• AxMultiServiceRouter src/ai/multiservice.ts: Model-based routing (route requests to specific providers by model name)

Sources: src/ai/wrap.ts69-181 src/ai/base.ts148-421 src/ai/types.ts src/ai/balance.ts40-299 src/ai/multiservice.ts src/ai/openai/api.ts src/ai/anthropic/api.ts src/ai/google-gemini/api.ts

Example: Complete Request Lifecycle

Here's what happens when you run this code:

Detailed Execution Sequence with Code Entities

Key Operations:

1. ai() Factory src/ai/wrap.ts300-346: Instantiates AxAIOpenAI provider
2. ax() Factory src/dsp/template.ts18-62: Parses signature string into AxSignature, creates AxGen
3. AxSignature.validate() src/dsp/sig.ts828-891: Ensures input has required review:string field
4. AxPromptTemplate.render() src/dsp/prompt.ts109-207: Generates prompt like:

   You are a classifier. Extract the following fields:
   - sentiment: class (one of: positive, negative, neutral)
   
   Input:
   review: "Great product!"
   

5. AxBaseAI.chat() src/ai/base.ts220-421: Rate limits, traces, delegates to OpenAI
6. extractValues() src/dsp/generate.ts1005-1193: Parses LLM response, validates against class enum

Sources: src/ai/wrap.ts300-346 src/dsp/template.ts18-62 src/dsp/parser.ts58-139 src/dsp/generate.ts77-1193 src/dsp/sig.ts828-891 src/dsp/prompt.ts109-207 src/ai/base.ts220-421 src/ai/openai/api.ts

Factory Functions

The framework provides four primary entry points for creating AI components:

Factory Function Overview

Factory	Returns	Purpose	Code Location
ai()	AxAI	Create AI service instances	src/ai/wrap.ts300-346
ax()	AxGen	Parse signatures into programs	src/dsp/template.ts549
agent()	AxAgent	Create multi-agent systems	src/prompts/agent.ts682-731
flow()	AxFlow	Build workflow orchestrations	src/flow/flow.ts582

Usage Patterns

Sources: src/ai/wrap.ts69-346 src/dsp/template.ts18-62 src/prompts/agent.ts682-731 src/flow/flow.ts40-582

Type Safety System

Ax provides compile-time and runtime type safety through its signature system:

Signature Definition to Runtime Validation

Type System Components

The type system is implemented across multiple layers:

Supported Field Types src/dsp/sig.ts17-31:

Type Category	Types	Parsed Representation
Primitives	string, number, boolean	Direct JSON values
Complex	json, class, array	Nested objects, enums, lists
Media	image, audio, file	Base64 data URIs or URLs
Temporal	datetime	ISO 8601 strings

Field Constraints and Modifiers:

• Optional fields: field?:type parsed by SignatureParser.parseField() src/dsp/parser.ts141-236
• Arrays: field[]:type creates AxFieldArrayType src/dsp/sig.ts32-39 with element type tracking
• Classes (enums): field:class "val1, val2, val3" with enumerated values stored in AxSignature._classes src/dsp/sig.ts237-265
• Descriptions: Inline via field:type "description text" or via fluent API f().input("field", f.string("desc"))

Three-Stage Validation Pipeline:

1. Parse-time validation: SignatureParser.validateFieldName() src/dsp/parser.ts309-344 enforces naming conventions (alphanumeric + underscore, starts with letter)
2. Input validation: AxSignature.validate() src/dsp/sig.ts828-891 checks required fields and types before execution
3. Output validation: validateAndParseFieldValue() src/dsp/util.ts13-190 parses LLM output and enforces type compliance

TypeScript Type Inference: Generic parameters AxSignature<IN, OUT> src/dsp/sig.ts300-364 enable full autocomplete for forward(ai, input: IN): Promise<OUT> via InferFieldValueType src/dsp/sig.ts41-96 utility type.

Sources: src/dsp/sig.ts17-891 src/dsp/parser.ts58-344 src/dsp/util.ts13-190

AI Provider Abstraction

The framework provides a unified interface for 15+ LLM providers through a three-layer architecture:

Provider Abstraction Layers

Provider Implementation Requirements

Each provider extending AxBaseAI src/ai/base.ts148-219 must implement three core methods:

Required Methods:

• createChatReq(req: AxChatRequest): ProviderRequest - Transform unified request to provider-specific format
• createChatResp(resp: ProviderResponse): AxChatResponse - Parse provider response to unified format
• createChatStreamResp(stream: SSEResponse): AsyncGenerator<AxChatResponse> - Handle streaming with SSE parsing via streamSSE() src/util/apicall.ts190-315

Provider-Specific Capabilities:

Feature	Implementation	Example Providers
Thinking tokens	thinking config property	AxAIGoogleGemini src/ai/google-gemini/api.ts270-320 AxAIAnthropic src/ai/anthropic/api.ts180-273
Schema cleanup	Remove unsupported JSON Schema properties	cleanSchema() in AxAIGoogleGemini src/ai/google-gemini/api.ts468-492
Multi-modal	Content part transformation	processContent() in AxAIOpenAI src/ai/openai/api.ts315-425
Function calling	Native tool support vs. prompt-based	AxAICohere uses tool_results src/ai/cohere/api.ts110-215
Embeddings	embed() method implementation	AxAIOpenAI.embed() src/ai/openai/api.ts513-594

Schema Constraints Example: Google Gemini does not support additionalProperties in JSON Schema, requiring schema transformation in createChatReq() src/ai/google-gemini/api.ts468-492

Sources: src/ai/base.ts148-421 src/ai/google-gemini/api.ts136-492 src/ai/anthropic/api.ts180-273 src/ai/openai/api.ts315-594 src/ai/cohere/api.ts110-215 src/util/apicall.ts190-315

Program Optimization

Ax provides three optimization algorithms for automatically improving AI program performance:

Optimizer Architecture

Optimizer Implementations

Optimizer	Algorithm	Implementation	Best For
AxMiPRO	Multi-objective Bayesian optimization (MIPROv2)	src/dsp/optimizers/miproV2.ts40-867	Multi-objective optimization with instruction and example tuning, no external dependencies
AxGEPA	Gradient Evolution + Pareto Analysis	src/dsp/optimizers/gepa.ts	Multi-objective optimization with Pareto-optimal solution discovery
AxGEPAFlow	Multi-module GEPA	src/dsp/optimizers/gepaFlow.ts	Complex workflows with multiple interconnected AxGen steps
AxACE	Agentic Context Engineering	src/dsp/optimizers/ace.ts1-848	Iterative playbook refinement via generator→reflector→curator loops
AxBootstrapFewShot	Trace-based demo generation	src/dsp/optimizers/bootstrapFewshot.ts15-180	Few-shot learning from successful execution traces
AxLearn	Unified optimization API	src/dsp/learn.ts9-211	Simplified interface replacing deprecated AxLearnAgent and AxTuner (v15.0.25+)

AxLearn Unified Interface: Introduced in version 15.0.25 CHANGELOG.md129-136 AxLearn provides a MIPROv2-like API that consolidates optimization capabilities previously split across multiple classes. It supports instruction optimization, demo bootstrapping, and parameter tuning through a single optimize() method src/dsp/learn.ts29-199

Sources: src/dsp/optimizers/miproV2.ts40-867 src/dsp/optimizers/gepa.ts src/dsp/optimizers/gepaFlow.ts src/dsp/optimizers/ace.ts1-848 src/dsp/optimizers/bootstrapFewshot.ts15-180 src/dsp/learn.ts9-211 CHANGELOG.md129-136

Common Optimizer Interface: All optimizers extend AxBaseOptimizer src/dsp/optimizer.ts400-650 providing:

• compile(options: AxCompileOptions): Promise<AxOptimizerResult> src/dsp/optimizer.ts651-723 - Main optimization loop with metric evaluation
• AxOptimizedProgram result src/dsp/optimizer.ts302-363 - Serializable configuration with toJSON() and applyOptimization()
• Checkpoint persistence via AxOptimizationCheckpoint src/dsp/common_types.ts185-203 with save/load callbacks
• Cost tracking through AxDefaultCostTracker src/dsp/optimizer.ts42-58 to enforce budget limits
• Progress logging via AxOptimizerLoggerFunction src/dsp/optimizerTypes.ts5-65

AxLearn Unified API: The AxLearn class src/dsp/learn.ts provides a simplified interface that replaced AxLearnAgent and AxTuner in version 15.0.25, offering a MIPROv2-like API for program optimization.

External Dependencies: All optimizers are pure TypeScript implementations with no external dependencies. Early versions of MiPRO required a Python service for Bayesian optimization, but this was eliminated in later releases.

Sources: src/dsp/optimizer.ts42-900 src/dsp/optimizers/miproV2.ts40-867 src/dsp/optimizers/gepa.ts src/dsp/optimizers/gepaFlow.ts src/dsp/optimizers/ace.ts1-848 src/dsp/optimizers/bootstrapFewshot.ts15-180 src/dsp/learn.ts9-211 src/dsp/common_types.ts185-203 src/dsp/optimizerTypes.ts5-65

Error Handling & Resilience

The framework implements multi-level error handling with automatic retry and correction mechanisms:

Error Handling & Resilience Architecture

Error Type Hierarchy [src/util/apicall.ts]:

Error Type	Retry Behavior	Use Case
AxAIServiceNetworkError	Automatic retry with exponential backoff	Network failures, DNS issues
AxAIServiceTimeoutError	Automatic retry with increased timeout	Slow responses, timeouts
AxAIServiceStatusError	Retry on 5xx, fail on 4xx	Server errors vs client errors
AxTokenLimitError	No retry (context too large)	Token limit exceeded
AxAIServiceAuthenticationError	No retry	Invalid API key
AxAIRefusalError	No retry	Content policy violation
AxAssertionError	Validation retry with correction	Output validation failures

Three-Level Retry Strategy:

1. Infrastructure Retry [src/util/apicall.ts]: Automatic retry for network/5xx/timeout errors

   • Exponential backoff: 1s, 2s, 4s, up to 60s with jitter
   • Respects Retry-After headers from rate limiting
   • Configurable via retryConfig on API options

2. Validation Retry src/dsp/generate.ts928-1003: Automatic retry for assertion failures

   • Adds correction message to conversation memory
   • Retries with full context (up to maxRetries times)
   • Useful for fixing formatting or constraint violations

3. Multi-Step Retry src/dsp/generate.ts196-240: Loops for function calling

   • Continues execution until no more function calls
   • Configurable maxSteps limit (default: 10)
   • Tracks step count and prevents infinite loops

Error Recovery Mechanisms:

• AxBalancer fallback [src/ai/balance.ts]: Automatically routes to backup providers on failure
• Cache invalidation: Errors prevent stale cached responses
• Detailed logging: Captures full request/response bodies for debugging
• OpenTelemetry events: Records errors in distributed traces

Sources: src/util/apicall.ts src/dsp/generate.ts196-1003 src/ai/balance.ts src/dsp/asserts.ts

Advanced Features

Context Caching

Context Caching System

Context caching reduces costs and latency by reusing expensive prompt prefixes (system prompts, examples, function definitions). Supported providers: Anthropic src/ai/anthropic/api.ts Google Gemini src/ai/google-gemini/api.ts

Cache Breakpoints:

• system: Cache system prompt only
• after-functions: Cache system prompt + function definitions
• after-examples: Cache system prompt + functions + few-shot examples

Sources: src/ai/anthropic/api.ts src/ai/google-gemini/api.ts CHANGELOG.md39-73

Streaming Support

Streaming Architecture

Streaming provides real-time response delivery with incremental value extraction. The framework maintains state across chunks and yields only new content deltas src/dsp/generate.ts699-926

Sources: src/dsp/generate.ts699-926 src/util/apicall.ts190-315 src/dsp/processResponse.ts

Memory Management

Memory System

AxMemory src/mem/memory.ts63-270 maintains conversational state across sessions with tag-based checkpointing for error recovery.

Sources: src/mem/memory.ts63-270 src/dsp/functions.ts205-457

Workflow Orchestration (AxFlow)

AxFlow src/flow/flow.ts40-580 provides declarative workflow construction with automatic dependency analysis and parallel execution:

AxFlow Features:

• Dependency analysis: AxFlowDependencyAnalyzer.analyze() src/flow/dependencyAnalyzer.ts17-129 builds dependency graphs
• Execution planning: AxFlowExecutionPlanner.plan() src/flow/executionPlanner.ts82-227 computes execution levels and parallelization
• Type-safe state: Generic TState parameter provides compile-time safety
• Control flow: branch(), while(), parallel(), feedback() primitives src/flow/flow.ts166-473
• Auto-parallelization: Independent nodes execute concurrently when enabled

Sources: src/flow/flow.ts40-582 src/flow/dependencyAnalyzer.ts17-129 src/flow/executionPlanner.ts82-227 src/flow/subContext.ts12-124

Multi-Agent Systems (AxAgent)

agent() factory src/prompts/agent.ts682-731 creates hierarchical agents with function calling:

Agent Features:

• ReAct execution: AxFunctionProcessor src/dsp/functions.ts205-457 handles tool loops
• Hierarchical composition: Child agents as tools via processChildAgentFunction() src/prompts/agent.ts59-141
• Smart model routing: addModelParameter() src/prompts/agent.ts545-590 enables model selection
• Field injection: Automatic input passing to child agents src/prompts/agent.ts79-114
• Function modes: native, prompt, auto execution src/dsp/functions.ts205-457

Sources: src/prompts/agent.ts59-731 src/dsp/functions.ts67-457 src/ai/base.ts196-218

Observability and Infrastructure

OpenTelemetry Integration

Comprehensive telemetry through OpenTelemetry:

Telemetry System

Metrics Instruments:

Instrument Category	Metrics	Implementation
Generation	Request count, latency (p50/p95/p99), retry count, assertion failures	AxGenMetricsInstruments src/dsp/metrics.ts48-76
AI Service	Token usage (input/output), cost estimates, request/response sizes, error rates	AxAIMetricsInstruments src/ai/metrics.ts14-43
Optimization	Compilation rounds, best score progression, convergence rate, evaluated configurations	AxOptimizerMetricsInstruments src/dsp/optimizer.ts60-91

Configuration via axGlobals: Set OpenTelemetry components globally via axGlobals src/dsp/globals.ts1-16:

State and Memory Management

Component	Purpose	Key Methods	Location
AxMemory	Multi-session conversation history	add(), addTag(), reset(), history()	src/mem/memory.ts63-270
AxDB	Vector database abstraction	upsert(), query()	src/db/wrap.ts11-77
AxDBManager	RAG with query rewriting and reranking	batchUpsert(), query(), search()	src/docs/manager.ts38-335
axGlobals	Global framework configuration	tracer, meter, logger, functionResultFormatter	src/dsp/globals.ts1-16

AxMemory Features:

• Session isolation: Each sessionId maintains independent conversation history
• Tag-based checkpointing: addTag(sessionId, tag) src/mem/memory.ts194-214 creates named checkpoints, resetToTag() reverts to saved state
• Streaming updates: updateMessage() src/mem/memory.ts161-171 allows incremental message modification during streaming responses
• History management: history(sessionId, count) src/mem/memory.ts216-243 retrieves conversation context with optional message limit

Vector Database Implementations:

Provider	Implementation	Key Methods
Pinecone	AxDBPinecone src/db/pinecone.ts	upsert(), query(), namespace-based partitioning
Weaviate	AxDBWeaviate src/db/weaviate.ts	upsert(), query(), schema management
Cloudflare	AxDBCloudflare src/db/cloudflare.ts	upsert(), query(), edge-deployed vectorization
In-Memory	AxDBMemory src/db/memory.ts	upsert(), query(), ephemeral storage for testing

All implementations conform to AxDBService interface src/db/types.ts with standardized upsert() and query() signatures.

External Integrations

Integration	Component	Purpose	Location
Vercel AI SDK	AxAIProvider	Bridge to Vercel AI SDK v5	src/aisdk/provider.ts
Apache Tika	AxApacheTika	Document format conversion (PDF, DOCX, etc.)	src/docs/tika.ts14-100
MCP Protocol	AxMCPClient	Model Context Protocol client for server tools	src/mcp/client.ts26-203
MCP Transports	AxMCPHTTPSSETransport, AxMCPStreambleHTTPTransport	HTTP/SSE transport implementations	src/mcp/transports/
Docker	AxDockerSession	Sandboxed code execution environment	src/funcs/docker.ts12-156

MCP Transport Options:

• HTTP SSE: AxMCPHTTPSSETransport src/mcp/transports/sseTransport.ts for server-sent events
• Streamable HTTP: AxMCPStreambleHTTPTransport src/mcp/transports/httpStreamTransport.ts for bidirectional streaming
• OAuth 2.1 support: AxMCPOAuthOptions src/mcp/oauth/types.ts with PKCE for secure authentication

AxDockerSession Usage: Provides sandboxed Python/Node.js code execution with configurable resource limits src/funcs/docker.ts12-156 Example:

Sources: src/aisdk/provider.ts src/docs/tika.ts14-100 src/mcp/client.ts26-203 src/mcp/transports/sseTransport.ts src/mcp/transports/httpStreamTransport.ts src/mcp/oauth/types.ts src/funcs/docker.ts12-156 src/dsp/generate.ts196-240 src/ai/base.ts220-421 src/dsp/globals.ts1-16 src/mem/memory.ts63-270 src/db/wrap.ts11-77 src/db/types.ts src/docs/manager.ts38-335

Version and Distribution

The framework follows semantic versioning with synchronized package versions:

• Current Version: 14.0.28 package.json76
• License: Apache-2.0 package.json9
• Node Requirement: >=20 package.json37-38
• Repository: https://github.com/ax-llm/ax package.json5-8

Build artifacts are generated using tsup and published to npm with provenance. The monorepo uses npm workspaces for managing inter-package dependencies.

Sources: package.json1-78 CHANGELOG.md1-204