Product
Understand the action. Keep the evidence.
Pramagent sits in the tool path of a connected agent. Your assistant stays where you already work.
Control my agent
Connect a supported tool path to apply policy checks before execution. Review the compatibility matrix, data handling and pilot terms before generating a credential. Installing a connector is optional when you only want to read evidence already in Pramagent.
Investigate existing activity
Open investigations for records already collected in your Pramagent workspace, or explore the no-login synthetic sample first. Reading existing evidence does not enable enforcement or replace your hooks. The private CLI supports bounded offline transcript/log import; that does not create a hosted upload journey or independently authenticate the original source. Ask the operator about distribution and keep existing controls in place.
Connect one tool path
Choose Claude Code for guided hook setup, or connect an OpenAI, Ollama, MCP, or custom application through its adapter. Saving a credential alone does not intercept tools. Direct calls outside the integration remain outside enforcement.
Choose the decision
Configure exact-tool Allow, Block, or Require approval rules. Built-in checks also apply. Claude can ask in its native permission prompt; other adapters currently use website review. Organization rules can require a separate reviewer.
Choose the evidence
Owners and admins can revisit Evidence settings to choose metadata-only, customer-retained inputs or operator-enrolled encrypted snapshots for new hosted coding decisions. Availability depends on operator capture policy and storage. This setting does not change enforcement, delete historical data or promise automatic retention periods.
Review the exact action
Website approvals bind to the proposed action and session and authorize one retry. Slack, Teams, and Telegram notifications link to the authenticated review page. Resume your agent after website approval.
Investigate the decision trail
Search session and policy metadata, inspect action fingerprints, and export a bounded investigation bundle. An allowed check does not establish that the tool executed.
Separate decisions from outcomes
A pre-tool check records what the policy allowed or denied. Supported post-tool hooks add a host-reported outcome receipt. Missing receipts stay visible and are never converted into claims that an action ran.