Roe Python SDK

A Python SDK for the Roe API.

v1.1.14 - SDK operation coverage is synchronized across Python, TypeScript, and Go. See SDK_EXAMPLES.md for copy-ready examples and use cases.

v1.0.0 - The SDK delegates transport and type details to the generated OpenAPI client behind stable public methods such as client.agents and client.policies. Noteworthy API and behavioral changes compared to earlier releases are listed in CHANGELOG.md.

Installation

uv add roe-ai

Quick Start

from roe import RoeClient

client = RoeClient(
    api_key="your-api-key",
    organization_id="your-org-uuid"
)

# Run an existing agent
job = client.agents.run(agent_id="agent-uuid", text="Analyze this text")
result = job.wait()

for output in result.outputs:
    print(f"{output.key}: {output.value}")

Or set environment variables:

export ROE_API_KEY="your-api-key"
export ROE_ORGANIZATION_ID="your-org-uuid"

Job Result Inspection

After waiting for a job, inspect its outcome using the JobStatus enum. The terminal status and error message are stuffed into the generated response's additional_properties and accessed via subscript:

from roe import JobStatus

result = job.wait()

# Check job outcome
if result["status"] in (JobStatus.SUCCESS, JobStatus.CACHED):
    for output in result.outputs:
        print(f"{output.key}: {output.value}")
elif result["status"] == JobStatus.CANCELLED:
    print("Job was cancelled")
elif result["status"] == JobStatus.FAILURE:
    print("Error:", result["error_message"])

# Available fields
result.outputs           # list[AgentDatum] (direct attribute on the generated model)
result["status"]         # JobStatus code (int) — set by Job.wait()
result["error_message"]  # Error string or None — set by Job.wait()

Errors

Non-2xx responses raise typed exceptions from roe.exceptions, all subclasses of RoeAPIException. Use them to handle expected failures without parsing error strings:

from roe.exceptions import (
    RoeAPIException,
    BadRequestError,            # 400 — validation / bad input
    AuthenticationError,        # 401 — missing or invalid API key
    InsufficientCreditsError,   # 402 — plan limit / billing
    ForbiddenError,             # 403 — feature or resource forbidden
    NotFoundError,              # 404 — resource not found
    ServerError,                # 5xx — server-side
)

try:
    client.agents.retrieve("00000000-0000-0000-0000-000000000000")
except NotFoundError as exc:
    print(exc.status_code, exc.message)

Every RoeAPIException also carries exc.headers (lowercase-keyed dict of the upstream response headers). Use it to read Retry-After on 429s or X-Request-Id for support tickets, without falling back to the raw httpx layer. Retry-After is preserved exactly as sent, so it may be numeric seconds or an HTTP-date:

try:
    client.agents.retrieve("00000000-0000-0000-0000-000000000000")
except RoeAPIException as exc:
    if exc.status_code == 429 and exc.headers:
        retry_after = exc.headers.get("retry-after")

job.wait() does not raise on agent-side failures — instead the returned result carries result["status"] == JobStatus.FAILURE and result["error_message"]. Transport / HTTP errors hit the typed hierarchy above.

SDK Operation Groups

Common operations are available directly on the SDK client.

engines = client.discovery.list_agent_engine_types()
models = client.discovery.list_supported_models(capability="text")

upload = client.tables.upload(
    table_name="customers",
    file="customers.csv",
    with_headers=True,
)

Agent Examples

Multimodal Extraction

Extract structured data from text and images:

agent = client.agents.create(
    name="Listing Analyzer",
    engine_class_id="MultimodalExtractionEngine",
    input_definitions=[
        {"key": "text", "data_type": "text/plain", "description": "Item description"},
    ],
    engine_config={
        "model": "gpt-5.5-2026-04-23",
        "text": "${text}",
        "instruction": "Analyze this product listing. Is it counterfeit?",
        "output_schema": {
            "type": "object",
            "properties": {
                "is_counterfeit": {"type": "boolean", "description": "Whether likely counterfeit"},
                "confidence": {"type": "number", "description": "Confidence score 0-1"},
                "reasoning": {"type": "string", "description": "Explanation"},
            }
        }
    }
)

job = client.agents.run(
    agent_id=str(agent.id),
    text="Authentic Louis Vuitton bag, brand new, $50"
)
result = job.wait()

Document Insights

Extract structured information from PDFs:

agent = client.agents.create(
    name="Resume Parser",
    engine_class_id="PDFExtractionEngine",
    input_definitions=[
        {"key": "pdf_files", "data_type": "application/pdf", "description": "Resume PDF"},
    ],
    engine_config={
        "model": "gpt-5.5-2026-04-23",
        "pdf_files": "${pdf_files}",
        "instructions": "Extract candidate information from this resume.",
        "output_schema": {
            "type": "object",
            "properties": {
                "name": {"type": "string"},
                "email": {"type": "string"},
                "skills": {"type": "array", "items": {"type": "string"}},
            }
        }
    }
)

job = client.agents.run(agent_id=str(agent.id), pdf_files="resume.pdf")
result = job.wait()

Web Insights

Extract data from websites with automatic screenshot/HTML/markdown capture:

agent = client.agents.create(
    name="Company Analyzer",
    engine_class_id="URLWebsiteExtractionEngine",
    input_definitions=[
        {"key": "url", "data_type": "text/plain", "description": "Website URL"},
    ],
    engine_config={
        "url": "${url}",
        "model": "gpt-5.5-2026-04-23",
        "instruction": "Extract company information from this website.",
        "vision_mode": False,
        "crawl_config": {
            "save_html": True,
            "save_markdown": True,
            "save_screenshot": True,
        },
        "output_schema": {
            "type": "object",
            "properties": {
                "company_name": {"type": "string"},
                "description": {"type": "string"},
                "products": {"type": "array", "items": {"type": "string"}},
            }
        }
    }
)

# Run the agent
job = client.agents.run(agent_id=str(agent.id), url="https://www.roe-ai.com/")
result = job.wait()

for output in result.outputs:
    print(f"{output.key}: {output.value}")

Interactive Web

Navigate websites and perform actions:

agent = client.agents.create(
    name="Meeting Booker",
    engine_class_id="InteractiveWebExtractionEngine",
    input_definitions=[
        {"key": "url", "data_type": "text/plain", "description": "Website URL"},
        {"key": "action", "data_type": "text/plain", "description": "Action to perform"},
    ],
    engine_config={
        "url": "${url}",
        "action": "${action}",
        "output_schema": {
            "type": "object",
            "properties": {
                "calendar_link": {"type": "string", "description": "Booking link found"},
                "steps_taken": {"type": "array", "items": {"type": "string"}},
            }
        }
    }
)

job = client.agents.run(
    agent_id=str(agent.id),
    url="https://www.roe-ai.com/",
    action="Find the founder's calendar link to book a meeting"
)
result = job.wait()

Rori Agents (Agentic Workflows)

Rori agents are autonomous investigation agents that follow policies (SOPs), use tools, and produce structured verdicts. Unlike extraction engines which transform data, Rori agents reason over evidence, apply policy rules, and return dispositions. All Rori agents are policy-aware — you define the rules, they run the investigation.

Policies

Policies define the rules, instructions, and disposition classifications that Rori agents follow. Creating a policy atomically creates the policy and its first version in one call:

policy = client.policies.create(
    name="AML Investigation Policy",
    content={
        "guidelines": {
            "categories": [
                {
                    "title": "Structuring",
                    "rules": [
                        {
                            "title": "Cash structuring below reporting thresholds",
                            "description": "Multiple deposits just under $10,000 within short timeframes",
                            "flag": "RED_FLAG",
                        }
                    ],
                },
                {
                    "title": "Layering",
                    "rules": [
                        {
                            "title": "Rapid movement between accounts",
                            "description": "Funds transferred through multiple accounts to obscure origin",
                            "flag": "RED_FLAG",
                            "sub_rules": [
                                {"title": "Cross-border wire transfers with no business purpose"},
                                {"title": "Shell company intermediaries"},
                            ],
                        }
                    ],
                },
            ]
        },
        "instructions": "Investigate the alert against each category. Use available data sources to gather evidence.",
        "dispositions": {
            "classifications": [
                {"name": "Suspicious", "description": "Activity warrants SAR filing"},
                {"name": "Not Suspicious", "description": "Activity has legitimate explanation"},
                {"name": "Needs Escalation", "description": "Requires senior analyst review"},
            ]
        },
        "summary_template": {
            "template": "Investigation of {{subject}} found {{verdict}} based on {{findings_count}} findings."
        },
    },
)

Iterate on policies by creating new versions:

# Create a new version (automatically becomes the current version)
new_version = client.policies.versions.create(
    policy_id=str(policy.id),
    content={...},  # Updated policy content
    version_name="v2 - added layering rules",
)

# List all versions
versions = client.policies.versions.list(policy_id=str(policy.id))

# Retrieve a specific version
version = client.policies.versions.retrieve(str(policy.id), str(new_version.id))

# Update policy metadata
client.policies.update(str(policy.id), name="Updated Policy Name")

# List all policies
policies = client.policies.list()

# Delete a policy
client.policies.delete(str(policy.id))

Policy Content Reference

Product Compliance

Analyze product listings against your compliance policy:

agent = client.agents.create(
    name="Product Compliance",
    engine_class_id="ProductPolicyEngine",
    input_definitions=[
        {"key": "product_listings", "data_type": "text/plain", "description": "Product listing to analyze"},
    ],
    engine_config={
        "policy_version_id": str(policy.current_version_id),
        "product_listings": "${product_listings}",
    },
)

result = client.agents.run_sync(
    agent_id=str(agent.id),
    product_listings="Nike Air Max 90, brand new, $45 — ships from Shenzhen",
)

AML Investigation

Investigate anti-money laundering alerts:

agent = client.agents.create(
    name="AML Investigation",
    engine_class_id="AMLInvestigationEngine",
    input_definitions=[
        {"key": "alert_data", "data_type": "text/plain", "description": "Alert data and context"},
    ],
    engine_config={
        "policy_version_id": str(policy.current_version_id),
        "alert_data": "${alert_data}",
    },
)

job = client.agents.run(
    agent_id=str(agent.id),
    alert_data="Customer John Doe, 5 cash deposits of $9,500 in 3 days",
)
result = job.wait()

Fraud Investigation

Investigate fraud alerts and suspicious activity:

agent = client.agents.create(
    name="Fraud Investigation",
    engine_class_id="FraudInvestigationEngine",
    input_definitions=[
        {"key": "alert_data", "data_type": "text/plain", "description": "Alert data and context"},
    ],
    engine_config={
        "policy_version_id": str(policy.current_version_id),
        "alert_data": "${alert_data}",
    },
)

job = client.agents.run(
    agent_id=str(agent.id),
    alert_data="Chargeback spike: 47 disputes in 24h from merchant ACME-1234",
)
result = job.wait()

Merchant Risk

Analyze merchant risk profiles:

agent = client.agents.create(
    name="Merchant Risk Analysis",
    engine_class_id="MerchantRiskEngine",
    input_definitions=[
        {"key": "merchant_context", "data_type": "text/plain", "description": "Merchant name and context"},
    ],
    engine_config={
        "policy_version_id": str(policy.current_version_id),
        "merchant_context": "${merchant_context}",
    },
)

job = client.agents.run(
    agent_id=str(agent.id),
    merchant_context="ACME Corp - Online electronics retailer, MCC 5732",
)
result = job.wait()

Agent Configuration Options

All Rori agents accept these options in engine_config:

Example with advanced configuration:

agent = client.agents.create(
    name="AML Investigation (Advanced)",
    engine_class_id="AMLInvestigationEngine",
    input_definitions=[
        {"key": "alert_data", "data_type": "text/plain", "description": "Alert data and context"},
    ],
    engine_config={
        "policy_version_id": str(policy.current_version_id),
        "alert_data": "${alert_data}",
        "reasoning_effort": "high",
        "context_sources": [
            {"type": "sql", "name": "Transactions DB", "connection_id": "conn-uuid"},
        ],
    },
)

Retry Behavior

Transient failures are retried with exponential backoff capped at about 10 seconds per attempt. By default there are up to 3 retries (configurable via max_retries or ROE_MAX_RETRIES):

client = RoeClient(
    api_key="your-api-key",
    organization_id="your-org-uuid",
    max_retries=5,  # default: 3
)

Retried: HTTP statuses 408, 429, and any 5xx, plus transport errors (for example disconnects and timeouts). JSON POST bodies may be replayed; multipart agent-run calls (run, run_sync, …) opt out via x-roe-skip-retry so they are not automatically retried at the transport layer.

Not retried immediately: Typical client/auth responses (401, 403, 404, validation 422, …) — surfaced as typed exceptions matching the SDK’s usual error mapping.

Batch Processing

When batch operations exceed 1,000 items, the SDK automatically chunks requests. A configurable delay (default: 10 seconds) is applied between chunks to avoid overwhelming the API. This applies to:

• client.agents.run_many() — job submissions
• client.agents.jobs.retrieve_status_many() — batch status checks
• client.agents.jobs.retrieve_result_many() — batch result retrieval

Single-chunk batches (≤1,000 items) are unaffected.

You can configure the delay via the batch_chunk_delay parameter or the ROE_BATCH_CHUNK_DELAY environment variable:

client = RoeClient(
    api_key="your-api-key",
    organization_id="your-org-uuid",
    batch_chunk_delay=2.0,  # default: 10.0
)

Running Agents

# Async (recommended)
job = client.agents.run(agent_id="uuid", text="input")
result = job.wait()

# Sync
outputs = client.agents.run_sync(agent_id="uuid", text="input")

# With files (auto-uploaded)
job = client.agents.run(agent_id="uuid", document="file.pdf")

# Batch processing
batch = client.agents.run_many(
    agent_id="uuid",
    batch_inputs=[{"text": "input1"}, {"text": "input2"}]
)
results = batch.wait()

# Skip the job-result cache and force a fresh run
# (available on run, run_sync, run_many, run_version, run_version_sync;
# the fresh result still refreshes the cache)
job = client.agents.run(agent_id="uuid", skip_cache=True, text="input")

Metadata

You can attach arbitrary metadata to any job when running an agent. Metadata is a dictionary of key-value pairs that gets stored with the job, useful for tracking, filtering, or correlating jobs with your own internal records.

# Attach metadata to an async job
job = client.agents.run(
    agent_id="agent-uuid",
    metadata={"customer_id": "cust-123", "request_source": "api"},
    url="https://example.com",
)
result = job.wait()

# Attach metadata to a sync job
outputs = client.agents.run_sync(
    agent_id="agent-uuid",
    metadata={"batch": "2026-02-12", "priority": "high"},
    url="https://example.com",
)

# Attach metadata to a batch of jobs (applied to all jobs in the batch)
batch = client.agents.run_many(
    agent_id="agent-uuid",
    batch_inputs=[{"url": "https://a.com"}, {"url": "https://b.com"}],
    metadata={"campaign": "weekly-scan"},
)
results = batch.wait()

# Attach metadata when running a specific version
job = client.agents.run_version(
    agent_id="agent-uuid",
    version_id="version-uuid",
    metadata={"experiment": "v2-prompt"},
    url="https://example.com",
)

# Reuse a retrieved agent id when building the run request
agent = client.agents.retrieve("agent-uuid")
job = client.agents.run(
    agent_id=str(agent.id),
    metadata={"source": "sdk"},
    url="https://example.com",
)

Agent Management

# List / Retrieve
agents = client.agents.list()
agent = client.agents.retrieve("uuid")

# Update / Delete
client.agents.update("uuid", name="New Name")
client.agents.delete("uuid")

# Duplicate
#
# Note: client.agents.duplicate(...) returns an AgentVersion (the new agent's
# first version), not a BaseAgent. The new agent's ID is reachable on the
# returned object as `.base_agent.id`:
duplicated_version = client.agents.duplicate("uuid")
new_agent_id = duplicated_version.base_agent.id

Version Management

# List and retrieve versions
versions = client.agents.versions.list("agent-uuid")
current = client.agents.versions.retrieve_current("agent-uuid")
version = client.agents.versions.retrieve("agent-uuid", "version-uuid")

# Create, update, delete versions
version = client.agents.versions.create(
    agent_id="agent-uuid",
    version_name="v2",
    input_definitions=[...],
    engine_config={...}
)

client.agents.versions.update("agent-uuid", "version-uuid", version_name="v2-updated")
client.agents.versions.delete("agent-uuid", "version-uuid")

# Run specific versions
job = client.agents.run_version("agent-uuid", "version-uuid", text="input")
result = job.wait()

Job Management

# Retrieve job status and results
status = client.agents.jobs.retrieve_status(job_id)
result = client.agents.jobs.retrieve_result(job_id)

# Batch operations
statuses = client.agents.jobs.retrieve_status_many([job_id1, job_id2])
results = client.agents.jobs.retrieve_result_many([job_id1, job_id2])

# Download references from jobs (screenshots, HTML, markdown)
content = client.agents.jobs.download_reference(job_id, resource_id)

# Delete job data
client.agents.jobs.delete_data(job_id)

Supported Models

Engine Classes

Development

Before opening a PR, format and lint the codebase by running:

./roe-cli format

CI runs the same checks (ruff check and ruff format --check) on every pull request and on merges to main, and they must pass before a PR can be merged.

Links

• Roe
• API Documentation
• Changelog
• Examples