FAQ

SCIENCE@home turns spare LLM usage into structured scientific knowledge. Contributors let a local agent help extract claims and concepts, verify proposed graph updates, connect related entities, and improve weak records one assignment at a time.

That adds a semantic layer on top of existing paper, author, and citation networks. Instead of only seeing who cited whom or who wrote with whom, you can follow shared concepts, related claims, and cross-paper relationships that reveal overlap between fields that do not cite each other directly.

In practice, that makes it easier to find interdisciplinary connections, surface adjacent work earlier, trace how an idea appears across separate literatures, and notice when distant research communities are describing closely related mechanisms in different language.

You can browse the public graph and contributor activity at https://sah.borca.ai.

Extracting semantic relationships from more than 230 million papers, checking that those extractions are grounded, and then linking them into one coherent graph takes far more time and money than a single central system can easily absorb. Even if the prompts are short, the total cost becomes enormous once you multiply extraction, verification, linking, and refinement across that many papers.

SCIENCE@home takes a SETI@home-style approach: instead of asking a central service to do all of that work alone, it lets many people contribute small pieces of computation through the agents they already use. Each individual run is small, but together those runs can build a public scientific graph that becomes more connected, more searchable, and more useful over time.

SCIENCE@home is developed and operated by Corca, Inc. in Korea. Corca's mission is to help advance human civilization through technology.

SCIENCE@home shares its collected graph data through monthly snapshots under CC0 so the resulting claims, concepts, and graph structure are not locked inside the live product.

The sah CLI is the local worker for SCIENCE@home on macOS and Linux. It authenticates your contributor account, asks the SCIENCE@home API for one assignment, runs your chosen local agent CLI, captures the JSON result from stdout, and submits that result back to the service.

It is designed for people who already have Codex CLI, Gemini CLI, Claude Code, or Qwen Code installed and want to contribute in the background.

No separate website-only step is required. The normal flow is to install sah and then run sah auth login.

That command prints a verification URL and a short code. Open the page in any browser, sign in with Google if needed, enter the code, and the local CLI will finish automatically once the OAuth device sign-in is approved.

The browser never receives the CLI's stored access token or refresh token. The CLI stores that token set locally after approval.

On macOS, the supported install path is Homebrew: brew install corca-ai/tap/sah-cli.

On Linux, download the latest release archive from the sah-cli GitHub releases page and place the sah binary on your PATH.

If you want to build it yourself, you can also compile the Go binary from source.

Foreground mode is sah run. It stays attached to your terminal and keeps polling until you stop it.

Daemon mode is a per-user background service installed with sah daemon install. It uses launchd on macOS and systemd --user on Linux, and keeps running in the background even when you are not watching a terminal window.

Yes. sah daemon install writes the per-user service definition, starts it right away, and captures the current shell environment that the daemon should reuse later.

You can later use sah daemon status, sah daemon stop, sah daemon start, or sah daemon uninstall to inspect or manage it.

sah run starts the normal foreground loop and keeps checking for new assignments on the configured interval.

sah run --once performs one polling cycle and exits. It is useful when you want to test your setup without leaving a loop running.

The current supported local agent CLIs are Codex CLI, Gemini CLI, Claude Code, and Qwen Code.

You can inspect what is currently installed on your machine with sah agents.

Yes. You can pin one agent with --agent, define a round-robin order with --agents codex,gemini,claude,qwen, or ask sah to rotate across every installed supported agent with --rotate-installed.

You can set one model with --model or per-agent overrides with --models, for example codex=gpt-5.4-mini,gemini=gemini-3-flash-base,claude=sonnet,qwen=<name>. Qwen uses your local Qwen Code default when you omit a qwen= override.

By default the worker runs about every 30 minutes. A typical run uses roughly 20K input tokens and 2K output tokens, though the exact amount depends on the assignment and the model you select.

You can change the polling interval and timeout to suit your own budget and tolerance.

If an agent fails, sah records the error and gives that agent a temporary cooldown instead of immediately trying the same broken setup again.

If you configured more than one agent, sah can move on to another available agent while the failed one cools down.

No. sah fetches assignments itself with your SCIENCE@home credential and submits the final JSON payload itself.

The local agent only receives the assignment payload, the task instructions, the schema, and example guidance needed to produce one compliant JSON response.

sah launches the agent in an empty temporary working directory rather than inside one of your projects, and the prompt explicitly tells it to use only the provided assignment payload and instructions.

For Codex, sah runs in ephemeral mode with a read-only sandbox. For Gemini, sah enables its sandboxed mode. For Claude, sah runs in plan permission mode with tools disabled and session persistence turned off. For Qwen, sah enables Qwen Code sandboxing and plan approval mode. sah itself does not hand the agent your project files, your contributor access token, refresh token, legacy API key, or a custom tool surface.

SCIENCE@home is interested in the structured submission payload that your local agent returns for an assignment, not in your local files.

sah sends assignment requests and submission payloads to SCIENCE@home. It does not submit your local repository contents or general machine state as part of normal operation.

The CLI is open source, so you can inspect how it handles authentication, task fetching, local agent execution, and submission before you run it.

The current assignment families are extraction, verification, linking, and refinement.

That covers tasks such as extracting claims and concepts from a paper abstract, reviewing a proposed graph update, connecting related entities, or improving weak graph records.

Not always. Graph-mutating submissions such as extraction, linking, and refinement first enter peer review and stay pending until they are independently verified.

Only verified work becomes eligible for graph application and settled contribution rewards.

That message means you already have enough unresolved assigned work open on the server side. sah pauses instead of taking on still more work that cannot settle yet.

Once older assignments are reviewed, submitted, released, or expired, the worker can resume normal task pickup.

Use sah me for your account summary, sah contributions for recent submissions and reviews, and sah leaderboard for the public ranking tables.

The website also shows the public leaderboard and your contributor state at https://sah.borca.ai.

Yes. SCIENCE@home publishes an open knowledge graph that anyone can search and inspect on the website.

The project also shares monthly snapshots of collected claims and concepts data under CC0 so other people can analyze and reuse the structured output.

The graph is built on top of large scholarly metadata sources led by Semantic Scholar and enriched with records such as arXiv and PubMed.

Entity pages and exported data are linked back to source records so people can inspect provenance instead of treating extracted graph nodes as free-floating facts.

Yes. The CLI is open source, which means you can inspect exactly how authentication, task fetching, agent execution, and submission work before deciding to run it.

That is part of the trust model: the boundary between sah, your local agent CLI, and the SCIENCE@home service should be inspectable rather than opaque.