synth-managed-research

Managed Research is Synth's product for applied AI teams that want repeatable, inspectable research workflows against real repos. Wave 1 is strongest at verification, eval execution, data assembly, and careful context optimization. The package exposes the SMR Python SDK and MCP server as the public control surface on top of the backend control plane.

Current status:

• SDK remigration is active in this repo
• MCP server modules live in managed_research.mcp
• standalone CLI migration is intentionally out of scope
• Data Factory and old onboarding or starting-data bootstrap APIs are intentionally out of scope

Project Notes vs Curated Knowledge

• project notes are durable notebook text for operator memory and local project context
• curated knowledge is the PG-backed durable store for org- or project-scoped findings you want later work to inherit

These are intentionally separate surfaces.

Python import surface:

from managed_research.sdk.client import SmrControlClient

Recommended launch flow:

• attach source repo or upload workspace files
• optionally set or append project notebook notes
• optionally set org or project curated knowledge
• check readiness
• preview lane with get_capacity_lane_preview
• inspect blockers with get_run_start_blockers
• trigger with trigger_run
• monitor progress
• retrieve the project workspace snapshot

Kickoff intent is queue-first:

• use initial_runtime_messages on blockers and trigger
• the legacy prompt field is no longer accepted
• migrate prompt="..." to initial_runtime_messages=[{"body": "...", "mode": "queue"}]

Project notebook, curated knowledge, and lifecycle helpers are separate:

• get_project_notes, set_project_notes, append_project_notes
• get_org_knowledge, set_org_knowledge, get_project_knowledge, set_project_knowledge
• pause_project, resume_project, archive_project, unarchive_project

For MCP callers, smr_trigger_run denials may come back as a successful tool result with top-level error, detail, message, and http_status. Always branch on result.get("error").

The package is structured as a library-first distribution rather than a standalone CLI product.

MCP

Primary path:

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 fallback:

uv tool install synth-managed-research
managed-research-mcp

The canonical MCP surface is owned by this package and is shared by the local stdio server and the hosted backend transport.