Stigmem OpenAI-Compatible Tool-Use Adapter

Package: stigmem-plugin-openai-tools-adapter

Exposes Stigmem's tools in OpenAI's function-calling format - compatible with LiteLLM, Ollama tool-use, and the OpenAI Python SDK. Use this when your OSS model deployment does not support MCP but does support the standard tools array in chat completion calls.

Design

The adapter has four surfaces:

1. tools() — returns STIGMEM_TOOLS, a list[dict] in OpenAI tool-use format. Pass directly to chat.completions.create(tools=...).

2. dispatch(tool_call) — executes a tool call returned by the model and returns an OpenAI-format {"role": "tool", ...} message dict. Accepts both SDK objects (ChatCompletionMessageToolCall) and plain dicts.

3. run_litellm(model, system_prompt, user_message) — thin agentic loop via LiteLLM. Works with any LiteLLM model string: "ollama/mistral", "openai/gpt-4o-mini", "anthropic/claude-3-5-haiku-20241022", etc.

4. run_openai(model, system_prompt, user_message, base_url, api_key) — loop via the OpenAI Python SDK. Set base_url="http://localhost:11434/v1" to target a local Ollama server.

The adapter depends on stigmem-node for plugin discovery metadata and the Python SDK client surface used by the adapter. litellm and openai are optional dependencies needed only for run_litellm() and run_openai() respectively.

Files

Installation

Requirements

• Python ≥ 3.11
• stigmem-py: pip install stigmem-py (or from workspace)
• Optional - for run_litellm(): pip install 'stigmem-plugin-openai-tools-adapter[litellm]'
• Optional - for run_openai(): pip install 'stigmem-plugin-openai-tools-adapter[openai]'
• Optional - local OSS models: install Ollama + pull a tool-capable model (ollama pull mistral)

Environment variables

STIGMEM_URL=http://localhost:8765
STIGMEM_API_KEY=sk-your-key                # optional
STIGMEM_SOURCE_ENTITY=agent:my-oss-agent  # entity URI for assertions

Enable

There is no node-global STIGMEM_*_ENABLED gate. Installing the package makes the plugin manifest discoverable through stigmem.plugins; a host application must import and call the adapter explicitly.

from stigmem_plugin_openai_tools import StigmemOpenAIToolsAdapter

adapter = StigmemOpenAIToolsAdapter.from_env()

Disable

Stop importing the adapter in the host application, remove any host-level model tool configuration that references STIGMEM_TOOLS, and uninstall the package when it is no longer needed.

Usage

Raw tool declarations (no LLM SDK needed)

from stigmem_plugin_openai_tools import STIGMEM_TOOLS

# Pass to any OpenAI-compatible endpoint
print(STIGMEM_TOOLS[0]["function"]["name"])  # → "assert_fact"

With LiteLLM — Ollama

from stigmem_plugin_openai_tools import StigmemOpenAIToolsAdapter

adapter = StigmemOpenAIToolsAdapter.from_env()

answer = adapter.run_litellm(
    model="ollama/mistral",
    system_prompt="You are a helpful agent with access to a shared knowledge base.",
    user_message="What role does user:alice have?",
)
print(answer)

With LiteLLM — OpenAI

answer = adapter.run_litellm(
    model="openai/gpt-4o-mini",
    system_prompt="...",
    user_message="...",
)

With OpenAI SDK against a local Ollama server

from stigmem_plugin_openai_tools import StigmemOpenAIToolsAdapter

adapter = StigmemOpenAIToolsAdapter.from_env()

answer = adapter.run_openai(
    model="mistral",
    system_prompt="You are a helpful agent.",
    user_message="Assert that user:alice has role engineer.",
    base_url="http://localhost:11434/v1",
    api_key="ollama",   # Ollama ignores this; SDK requires a non-empty value
)

Manual dispatch loop

import litellm
from stigmem_plugin_openai_tools import StigmemOpenAIToolsAdapter

adapter = StigmemOpenAIToolsAdapter.from_env()
messages = [
    {"role": "system", "content": "You are an agent with Stigmem access."},
    {"role": "user", "content": "What facts exist about project:loom?"},
]

response = litellm.completion(
    model="ollama/mistral",
    messages=messages,
    tools=adapter.tools(),
)

msg = response.choices[0].message
messages.append(msg.model_dump())

for tc in msg.tool_calls or []:
    messages.append(adapter.dispatch(tc))

# Continue the loop until no more tool calls...

Test

cd stigmem
uv run pytest experimental/openai-tools-adapter/tests -q

No live node, Ollama, or LiteLLM key required — all HTTP calls are mocked with respx.

Uninstall

python -m pip uninstall stigmem-plugin-openai-tools-adapter

Ollama model compatibility

Tool use requires a model that supports function calling. Tested models:

Pull a tool-capable model: ollama pull mistral

Protocol notes

• The OpenAI format uses lower-case JSON Schema type strings ("string", "object", "array"). This is the opposite of Gemini's upper-case format.
• dispatch() accepts both SDK model objects and plain dicts, making it easy to use from any framework or in tests.
• The run_litellm() and run_openai() loops cap at max_rounds=10.
• Dispatch errors are returned as {"error": "..."} in the tool message so the model can reason about failures.

Package Lifecycle

stigmem-plugin-openai-tools-adapter is an experimental, opt-in adapter package for the v0.9.0a10 adapter publication batch. Installing the package makes it discoverable through the stigmem.plugins entry-point group. The adapter performs no node-global work unless a host application imports it and calls its surfaces.