quartermaster-sdk 0.3.1

Quartermaster — modular AI agent orchestration framework. Install this to get all packages.

Quartermaster SDK

Modular AI agent orchestration framework by MindMade.

Quartermaster lets you build AI agent workflows as directed graphs — define nodes (LLM calls, decisions, user input, tools), connect them with edges, and execute them with a pluggable engine.

Quick Install

# Core framework (graph + providers + tools + nodes + engine)
pip install quartermaster-sdk

# With OpenAI
pip install quartermaster-sdk[openai]

# With everything (all providers, all tools, MCP client, code runner)
pip install quartermaster-sdk[all]

Quick Start (local Ollama, zero config)

ollama pull gemma4:26b      # or any model you've pulled

import quartermaster_sdk as qm

qm.configure(
    provider="ollama",
    base_url="http://localhost:11434",   # or set $OLLAMA_HOST
    default_model="gemma4:26b",
)

# Graph() auto-creates Start; .end() / .build() are both optional when running via qm.run().
result = qm.run(qm.Graph("chat").user().agent(), "Pozdravljen, koliko je ura?")
print(result.text)

Single-shot helpers (no graph visible)

# prompt → str
reply = qm.instruction(system="Respond in Slovenian.", user="Pozdravljen!")

# prompt → Pydantic model (typed JSON extraction)
from pydantic import BaseModel

class Classification(BaseModel):
    category: str
    priority: str

data = qm.instruction_form(Classification, system="Classify.", user=email_body)

Reading specific node outputs with capture_as=

graph = (
    qm.Graph("enrich")
    .agent("Research", tools=[...], capture_as="notes")
    .instruction_form(CustomerData, system="Extract.", capture_as="data")
)
result = qm.run(graph, "VT-Treyd Slovenija")
result["notes"].output_text    # agent's free-text research
result["data"].output_text     # extracted JSON

Streaming (v0.3.0 filtered iterators)

qm.run.stream(...) returns a wrapper you can iterate raw or pipe through a filter — one helper per chunk family:

Filter	Yields	Use for
.tokens()	str (the token text)	Typewriter UI — just the text
.tool_calls()	ToolCallChunk	Dashboard cards: call.tool, call.args
.progress()	ProgressChunk	prog.message, prog.percent, prog.data
.custom(name=...)	CustomChunk	Application-defined milestones
(raw for chunk in ...)	Chunk union	Debugging, pass-through consumers

# Typewriter effect -- tokens only.
for token in qm.run.stream(graph, "Tell me a story").tokens():
    print(token, end="", flush=True)

# Dashboard view -- just the tool calls.
for call in qm.run.stream(graph, "Research Slovenia").tool_calls():
    ui.tool_card(call.tool, call.args)

# Progress cards interleaved with model tokens.
for prog in qm.run.stream(graph, "Crunch the dataset").progress():
    ui.status(prog.message, prog.percent)

# Subscribe to one milestone name only.
for evt in qm.run.stream(graph, "Research").custom(name="source_found"):
    ui.add_source(evt.payload["url"])

Streams are single-pass — the wrapper owns its underlying generator, so picking a second filter (or raw-iterating after a filter) raises RuntimeError("stream already consumed"). Pick one consumer per stream.

The async analogue is available via qm.arun.stream(...) with the same four filter helpers, returning AsyncIterator[...].

Post-mortem Result.trace

Every Result (sync or the terminal DoneChunk.result of a stream) carries a structured Trace built from the full FlowEvent stream:

result = qm.run(graph, "Hello!")

result.trace.text                        # concatenated model output
result.trace.tool_calls                  # list[dict] across every agent node
result.trace.progress                    # list[ProgressEvent]
result.trace.custom(name="source_found") # filtered CustomEvent list
result.trace.by_node["Researcher"].text  # tokens for a single node
print(result.trace.as_jsonl())           # JSONL export for logs / fixtures

Progress events from inside tools

Long-running tools reach the flow's ExecutionContext via qm.current_context() and emit structured events that stream back to the UI alongside model tokens:

from quartermaster_tools import tool

@tool()
def slow_research(topic: str) -> dict:
    ctx = qm.current_context()      # None when called outside a flow -- safe
    if ctx is not None:
        ctx.emit_progress("Gathering sources", percent=0.25, topic=topic)
        ctx.emit_custom("source_found", {"url": "https://example.com"})
    # ... do real work ...
    return {"summary": "..."}

OpenTelemetry instrumentation

pip install 'quartermaster-sdk[telemetry]'

from quartermaster_sdk import telemetry

telemetry.instrument()     # uses the global tracer provider
qm.run(graph, "Hello!")    # every node + tool call is now a span

Spans follow the OpenTelemetry GenAI semantic conventions (gen_ai.system, gen_ai.operation.name, gen_ai.tool.name, gen_ai.usage.input_tokens, …). Point your exporter at Jaeger, Tempo, Honeycomb, Logfire, Phoenix, or any OTLP collector.

Quick Start (cloud provider)

agent = (
    qm.Graph("My Agent")
    .user("What can I help you with?")
    .instruction("Respond", model="gpt-4o", system_instruction="You are a helpful assistant.")
)
result = qm.run(agent, "How does photosynthesis work?")

Sync chat shim (no graph needed)

For one-shot LLM calls from sync code (Celery workers, Django views, CLI scripts) — no asgiref.async_to_sync wrapper required:

from quartermaster_providers.providers.local import OllamaProvider

provider = OllamaProvider(default_model="gemma4:26b")
result = provider.chat(
    messages=[{"role": "user", "content": "Pozdravljen!"}],
    max_output_tokens=128,
    thinking_level="off",
)
print(result.content)        # promoted from `reasoning` if `content` is empty
print(result.usage)          # {prompt_tokens, completion_tokens, total_tokens}

Packages

Package	Description
quartermaster-graph	Graph schema, builder API, validation
quartermaster-providers	LLM provider abstraction (OpenAI, Anthropic, Google, Groq, local)
quartermaster-tools	Tool definition, registry, built-in tools
quartermaster-nodes	Node execution protocols and implementations
quartermaster-engine	Flow execution, traversal, memory, streaming
quartermaster-mcp-client	MCP protocol client (standalone)
quartermaster-code-runner	Docker sandboxed code execution (standalone)

Documentation

See the docs/ directory:

• Getting Started
• Graph Building
• Architecture
• Tools Catalog
• Providers
• Security

License

Apache 2.0