The Client API

The constructor

If you have already built an AutoGen-compatible model client, construct MetadataReviewerClient directly:

MetadataReviewerClient(
    model_client,       # a pre-built AutoGen ChatCompletionClient
    assets_dir=None,    # directory holding agent-manifest YAML files
)

Parameter	Description
model_client	Any object implementing AutoGen's ChatCompletionClient protocol. The factory methods below build this for you.
assets_dir	Optional path to a directory of manifest YAML files. Defaults to the agents_manifest/ directory bundled inside the package. Set this to point at your own custom manifests.

assets_dir vs. manifest_file

The directory of manifests is set once, on the constructor via assets_dir. The specific file to use is chosen per submission via the manifest_file argument to submit(). Do not pass assets_dir to submit() — it is not a parameter there.

Factory classmethods

These four classmethods cover the common providers. All four also accept the optional assets_dir argument.

from_openai

Parameter	Description
model	Model name, e.g. "gpt-4o", "gpt-4o-mini".
api_key	Your OpenAI API key.

For reproducibility the client is built with a fixed seed=1029 and temperature=0 (or 1 for gpt-5 models, which require it). It also advertises JSON / structured-output support to AutoGen.

from_anthropic

Parameter	Description
model	Model name, e.g. "claude-sonnet-4-6", "claude-haiku-4-5".
api_key	Your Anthropic API key.

from_azure

Parameter	Description
model	Model name, e.g. "gpt-4o".
azure_endpoint	Azure OpenAI endpoint URL.
azure_deployment	Azure deployment name.
api_version	API version string, e.g. "2024-02-01".
azure_ad_token_provider	Optional token-provider callable from azure.identity.
azure_ad_token	Optional static Azure AD token; use when no provider is available.

from_ollama

Parameter	Description
model	Model name, e.g. "llama3.2", "mistral".
host	Host of the Ollama server. Defaults to "http://localhost".
port	Port of the Ollama server. Defaults to 11434.

Submitting jobs: submit vs. submit_async

Both methods accept the same arguments and return a Job immediately. They differ only in how the pipeline is scheduled.

Argument	Description
metadata	The metadata to scan, as a dict or a JSON string.
manifest_file	Name of the YAML manifest file inside assets_dir. Defaults to the bundled default_agents_manifest.yml.
team_preset	AutoGen team routing strategy (see Advanced Usage). Defaults to "RoundRobinGroupChat".

Method	Use when
submit(...)	You are in ordinary synchronous code (a script, a REPL, a notebook). The pipeline runs in a daemon thread with its own event loop, so it is safe to call even when no event loop exists.
await submit_async(...)	You are already inside an async context. It schedules the pipeline as an asyncio Task in the current event loop.

Synchronous example:

job = client.submit(metadata, team_preset="RoundRobinGroupChat")
result = job.wait_sync(timeout=300)

Asynchronous example:

async def review(metadata):
    job = await client.submit_async(metadata)
    return await job.wait(timeout=300)

Job-management methods

The client tracks every job it creates so you can look them up later.

Method	Description
get_job(job_id)	Return the Job with the given ID, or raise KeyError.
list_jobs()	Return all tracked jobs.
cleanup_jobs(keep_statuses=None)	Remove finished jobs from the registry and return how many were removed. By default it keeps only pending and running jobs.
list_manifests()	Return the available YAML manifest file names found in assets_dir.

# discard everything that has finished, keeping only active jobs
removed = client.cleanup_jobs()
print(f"Cleared {removed} finished jobs")

# which manifests are available?
print(client.list_manifests())  # ['default_agents_manifest.yml', ...]