managed-research 1.2026.50601

SDK, MCP, and schemas for authenticated Managed Research.

Managed Research

Hosted AI workers for durable repo and research runs.

Documentation: https://docs.usesynth.ai/managed-research/intro

Use Managed Research when you want a repeatable run instead of an untraceable chat transcript: attach repositories and context, launch work from Python or MCP, and read back logs, checkpoints, artifacts, PRs, usage, and final reports.

Engineering style

Contributors should follow Synth Style (engineering rules for contracts, errors, and structure). The canonical Markdown copy lives in the backend repository: specifications/tanha/references/synthstyle.md.

Surfaces

This package is for authenticated Managed Research integrations. The Python SDK and MCP server expose the backend catalog for project and run automation, including the runbook presets, work modes, model catalog, and usage state the authenticated org is entitled to use.

Surfaces and tiers

• Open Research web: public lab surface with lite and open_ended_discovery.
• Authenticated Free: lite, directed_effort, standard model access, and one-time Managed Research credits.
• Authenticated Pro and Team: Free plus the paid GA model bundle and higher Managed Research usage pools.
• Beta Access: an overlay on Free, Pro, or Team that adds open_ended_discovery, heavy, and all-model access.

Open Research on the public web is a narrower lab surface: it accepts lite runs with open_ended_discovery only. Treat those limits as web admission policy, not as SDK or MCP client-side rules. See the customer docs for the current product split: https://docs.usesynth.ai/managed-research/intro

Installation

uv add managed-research

Set an API key:

export SYNTH_API_KEY="sk_..."

Quickstart

import os

from managed_research import ManagedResearchClient, ProjectSelector

client = ManagedResearchClient(api_key=os.environ["SYNTH_API_KEY"])

run = client.runs.start(
    "Review the project context and propose the smallest high-impact improvement.",
    host_kind="daytona",
    work_mode="directed_effort",
    providers=[{"provider": "openrouter"}],
)

print("run:", run.run_id)

result = run.wait(timeout=60 * 60, poll_interval=15)
print("state:", result.state.value)
print("artifacts:", [artifact.title for artifact in run.artifacts()])

Main Ideas

• Use ManagedResearchClient as the Python entrypoint.
• Use client.runs.start(...) for runs that default to the caller's Miscellaneous project.
• Pass project_id= or project=ProjectSelector.from_project_id(...) when a run belongs to a specific existing project.
• Use client.projects.create(...) and client.project(project_id) for durable project-scoped work.
• Run preflight before launch when you want structured launch blockers.
• Inspect runs through messages, timelines, event logs, authority readouts, operator evidence, traces, checkpoints, artifacts, usage, questions, approvals, and actor/task counts.
• Use MCP from Codex or Claude Code when you want an agent-native interface.

Resource limits

Managed Research exposes resource limits as first-class run and project state. Use resource_limits() to inspect configured caps and progress_toward_resource_limits() to inspect current usage, active blockers, and extension posture.

When a spend limit blocks work, call request_resource_limit_extension(...) from a run, project, or ManagedResearchClient. Extensions accept either a new absolute USD limit or an added USD amount, plus optional blocker resolution and resume behavior. MCP exposes the same flow through smr_get_run_resource_limits, smr_get_run_progress_toward_resource_limits, smr_get_project_resource_limits, smr_get_project_progress_toward_resource_limits, and smr_request_resource_limit_extension.

MCP

Hosted:

codex mcp add managed-research --url https://api.usesynth.ai/mcp
claude mcp add --transport http managed-research https://api.usesynth.ai/mcp

Local stdio:

uv tool install managed-research
managed-research-mcp

Links

• Quickstart
• MCP quickstart
• Python SDK quickstart
• Launch fields
• Models and harnesses
• Runs and evidence