Projects and Context

Use projects to bind repos, files, datasets, credentials, notes, and reusable knowledge.

Projects keep the context that workers need across runs.

When to use a project

Use a project when work depends on:

• a GitHub repo or repo set
• uploaded files or starting data
• datasets, traces, papers, or benchmark material
• workspace credentials and provider bindings
• project notes or durable knowledge
• recurring budgets and run policy

Use a one-off run when the goal is self-contained and the default project is enough.

Project request bodies

The released project client validates the complete runnable-project mapping before making a network request. Supply a pool and agent profiles from your organization setup:

project_request = {
    "name": "Improve eval reliability",
    "timezone": "UTC",
    "pool_id": "pool_...",
    "runtime_kind": "sandbox_agent",
    "environment_kind": "harbor",
    "agent_profiles": {
        "orchestrator_profile_id": "profile_...",
        "default_worker_profile_id": "profile_...",
        "worker_profile_ids": [],
    },
}
created = client.projects.create_runnable(project_request)

Replace the placeholder IDs with resources available to your organization. Preflight checks whether that configured project is launch-ready; it does not fill missing creation fields. Keeping the request as a mapping avoids importing an internal or unreleased project-spec alias.

Typical setup

created = client.projects.create_runnable(project_request)
project = client.project(created.project_id)
project.repositories.attach(github_repo="owner/repo")
project.context.set_project_knowledge(
    "Prioritize deterministic evals, clear logs, and small reviewable patches."
)

Then preflight before launching:

preflight = project.runs.preflight(
    host_kind="daytona",
    work_mode="directed_effort",
    providers=[{"provider": "openrouter"}],
    runbook="lite",
)

Context quality

Good project context gives workers:

• the target repo and paths
• the command that proves success
• known failure modes
• budget and time constraints
• review criteria
• which files or APIs are off limits

Keep credentials out of prose. Attach them through supported credential and provider mechanisms.

Onboarding blockers

Some projects cannot launch until setup is complete. Preflight can report blockers for repo access, missing credentials, missing runtime capacity, provider availability, budget caps, or incomplete project setup.

Using projects in the web app (logged in)

When you are signed in, Managed Research projects open as a guided workspace in the browser. The UI does not mirror every MCP tool, but it covers launch, inspection, steering, and results for lite and heavy presets and directed effort and open-ended discovery work modes. That is a wider launch surface than the public Open Research lab, which is intentionally narrow on the open web.

Use these routes as a map (replace [projectId] and [runId] with your IDs):

Area	Typical route	What you do there
Project hub	/smr/[projectId]	Dashboard, run summaries, entry to the active run.
Onboarding	/smr/[projectId]/onboarding	Complete setup so preflight and launches are unblocked.
New run	Launch UX (from project)	Choose preset, work mode, models, files, and kickoff; quick-create paths for miscellaneous projects.
Runs list	/smr/[projectId]/runs	Browse and open runs.
Run console	/smr/[projectId]/runs/[runId]	Timeline, live ops, workers, runtime stream and transcript, messaging, checkpoints, branch, stop, pause, resume.
Approvals and questions	Routes under project + run	Human approvals and run questions.
Results	/smr/[projectId]/results/*	Pull requests and durable outputs.
Resources	/smr/[projectId]/resources/*	Files, datasets, repositories, connections, knowledge and context.
Usage and settings	/smr/[projectId]/usage, /smr/[projectId]/settings/*	Usage views and project settings (including exports).
Global	/smr/new, /smr/projectsList	Create projects and browse the list.
Artifact link	/smr/artifacts/[artifactId]	Deep link to an artifact when you have its identifier.

For automation (agents, scripts, CI), prefer MCP or SynthClient().research; the web app is best for interactive review and steering. See Get started, Python SDK guide, and the generated API reference.