# Install Cemented research

Install the complete Cemented research bundle (version 0.9.7). It contains
three sibling skills: research, search, and audit, with references,
examples, and Python helpers for direct API requests and offline preparation.
Downloading only SKILL.md is incomplete.

For app-specific setup, visit /connect on this site's origin. The human-readable
skill guide is /connect/skills, plugin guide /connect/plugins, and MCP guide /connect/mcp.

## Install the plugin in Claude

On a paid Claude plan, download /connect/plugins/download from this site's origin.
In Claude web, open Customize → Plugins and upload the custom plugin ZIP.
Open the installed Cemented plugin → Connectors → Connect beside cemented, then
complete the Cemented OAuth sign-in. Uploading the ZIP does not complete this
step; Claude may show “Not added” without prompting you in chat.
Start a new chat, enable Cemented under + → Connectors, and ask for a private
Cemented research report. If the plugin's connector is absent, add
https://www.cemented.ai/mcp through Customize → Connectors → + → Add custom
connector. Use the plugin's existing connector when present.
No terminal is needed. Organization settings may restrict installation.

The plugin bundles all three skills and the MCP connection. The manifest's plugin
entry provides its exact archive path, version, size, and SHA-256 checksum.
ChatGPT public installation requires a separate OpenAI directory submission;
uploading this ZIP as a chat attachment does not install a ChatGPT plugin.

## Connect your agent

If CEMENTED_API_TOKEN is configured in your agent's environment, use it for direct
API requests. The bundle's search/scripts/cemented.py helper uses Python 3 with no
third-party dependencies. Its connection command reports api or mcp without
printing credentials or making a request. It does not load .env files. Configure
the token in your host's secret settings or environment before starting the agent;
do not put it in the bundle, chat, or command-line arguments. You can create your
own token at https://www.cemented.ai/settings/tokens. CEMENTED_API_URL optionally
selects the origin for your requested environment; it defaults to
https://www.cemented.ai. API calls use your token's account and retention rules;
zero-data-retention mode does not support report uploads. An API error must not
silently switch accounts or retry through MCP.

Without an API token, add https://www.cemented.ai/mcp in your agent's connection
settings. It uses Streamable HTTP and OAuth. Let the user finish Cemented's
sign-in and connection flow. The host stores credentials. Use another endpoint only
when the user explicitly requests that environment.

An agent with remote MCP and OAuth support can search, fetch sources, validate,
upload, read, and open reports through the server. Usage is billed to the connected
Cemented account with standard retention. If the host cannot install local skills,
connect MCP directly; do not claim a SKILL.md file has been installed there.

## Download and install the research workflow

No Cemented login is needed to download the bundle. Local skill support is required
for installation. Python 3 runs the bundled API helper and optional offline
helpers. Hosts with their own HTTP tools can call the same API directly; web
hosts without an execution environment can use MCP.

1. Use the origin of this document's URL as BASE. Fetch BASE/connect/skills/manifest.json.
   It provides the current version, archive path, SHA-256 checksum, and skill names.
2. Download BASE plus the manifest's archive path into a temporary directory.
   Verify the SHA-256 of the downloaded bytes against the manifest before extracting.
3. Inspect the archive and extract it into that temporary directory. Reject absolute
   paths, parent traversal, and symlinks. The skill directories are under .agents/skills/.
4. Install research, search, and audit as sibling directories in the
   current agent's supported skill location. Use the user's requested scope, or the
   current project by default. Follow that host's installation instructions. Preserve
   complete directory contents; do not create duplicates in multiple locations or
   silently overwrite an existing modified skill.
5. If Python is available, run python3 <skill-directory>/research/scripts/publish.py --help
   and python3 <skill-directory>/search/scripts/search.py --help. Report the
   installed version and location. Reload skills if your host requires it.

## First research request

Read research/SKILL.md and search/references/connection.md. Select API requests
when a token is configured; otherwise discover the Cemented MCP tools. Collect
source text with its URL-bound contentSha256, prepare selected
evidence, and follow the research and writer feedback contract. Call report_schema
for the current artifact format, then validate_report and upload_report. Keep the
same idempotency_key for retries. Reports are private by default; publish only
when the user explicitly asks for a public link.

Open the report promptly in the host's in-app browser/sidebar when supported.
API uploads return url, which may require browser sign-in for a private report.
MCP uploads also return open_url for temporary access without sign-in; call MCP
open_report for a fresh link if needed. Do not switch from API to MCP to open a
report under another identity. Verify the report
renders before saying it is open. Return the durable url in your final response;
do not put the temporary access link in report content or shared messages. Without
an in-app browser, return the durable private URL and explain it may need sign-in.

## Updates

Fetch the manifest when the user asks to update. Compare version and checksum,
review local modifications, and replace all three skill directories together.
The host's credentials and MCP connection are managed separately from skill files.