SandBase Skills

Use multi-source-search to turn a research question into source-linked findings, conflicts and evidence gaps. It uses search and page-reading tools already available in a compatible agent; SandBase providers are optional for this skill. The repository also includes specialized social, marketing and business workflows, some of which depend on SandBase capabilities. These are agent instructions, not a standalone search engine or a single MCP server. Read the core workflow.

Good fit: an agent user who wants a defined research process and a structured evidence ledger. Choose another tool if you need a turnkey hosted research service, guaranteed source accuracy, or proof that every included workflow has been tested. We reviewed selected source paths; we did not run the whole collection.

The reviewed snapshot contains 245 source SKILL.md files across research and marketing. GitHub’s short description still says 88, so use the versioned catalog and individual skill requirements when choosing what to install.

The repository has an Apache-2.0 root license and third-party notices. Downloading skill instructions does not include free access to an AI model, search provider or hosted API.

• Core research workflow: multi-source-search does not require a SandBase account when the host already provides web search and page reading. Your host’s subscription, model and search charges still apply. Requirements.
• Specialized SandBase workflows: an account and API key may be needed. The official pricing page describes usage-based billing per call, with rates that vary by model or API. Check the specific capability’s price before running it; do not assume every call is free.
• Limited free allowance: the pricing page currently advertises 100 eligible data-API calls per organization per day, resetting at 00:00 UTC. This is not a blanket allowance for every model or workflow. We have not tested account eligibility or billing. Current pricing and conditions.

This directory listing is free and does not require a backlink. It does not include API credit, a paid account or a favorable security rating.

Before installing, read the multi-source-search instructions and bundled files. You need a compatible agent with search and page-reading tools for live research, plus Python 3.9 or later only if you want to run the offline validator.

The upstream README documents this standard Agent Skills installation command:

npx skills add sandbaseai/sandbase-skills@multi-source-search

This is the documented single-skill route, not an installation we have tested. The command can download and execute the third-party skills CLI and install skill instructions. Review its prompts, choose the intended agent and project, and inspect the resulting files before enabling the skill. The command follows upstream content; it does not pin the reviewed commit. Upstream installation documentation.

Start with a public, non-sensitive question. Check that your agent actually has search and page-reading tools, and inspect the citations in its output. Do not connect an API account just to compensate for a missing host capability without checking the data and cost implications.

At reviewed commit cbab581, the repository’s own DSH CLI drops the first positional argument when --project is omitted. We reproduced this with the read-only list command: it printed a usage error instead of listing skills. The same parser affects the documented add form without an explicit project. Affected command parser.

This is a different installer from the standard npx skills add route above. We are not recommending a DSH whole-bundle installation or claiming that an alternative command is tested. If you need DSH, check the latest upstream instructions for a fix before proceeding.

The installer also supports --force and removal operations that delete an existing skill directory. Back up custom instructions and avoid blind retries with overwrite flags. Filesystem implementation.

After checking out and inspecting the reviewed repository, its committed example can be checked locally:

python research/multi-source-search/scripts/validate_report.py examples/verifiable-research-report.json

Our local check on November 9, 2026 returned:

VALID: 3 source(s), 1 claim(s), 2 provider(s)

The validator reads local JSON and checks its structure, source references, provider diversity and declared confidence. Passing means internal consistency; it does not prove that the cited sources are true or genuinely independent. Example input.

For your own report, inspect the actual source pages as well as the validator result. If a provider fails or evidence is missing, preserve that gap rather than retrying paid calls automatically. Never put API keys into a prompt or report; sending sensitive queries or documents to a provider requires a separate decision.

This entry is based on snapshot cbab581: repository metadata, licensing notices, installation code, the core research skill and validator, plus selected workflow instructions. We reproduced the DSH list failure and validated the supplied offline example. We did not install the whole pack, verify every skill’s behavior, test billing, or independently establish the rights to every upstream source.

The repository is accepted as a normal editorial listing. No security grade, verified badge, production-safety guarantee or ranking benefit is awarded. Check the review methodology for the distinction between source review and a security audit.