Skip to content

Tools and effects

Language models should not directly perform destructive side effects. When a model attempts to write a file, execute a shell script, or send an email, Seepient splits the process into two phases: preparation and brokered execution.

Model generates a tool call

Tool creates a PreparedActionDraft

Permission Engine checks policy & consent mode

Approval Broker confirms the action

Execution Boundary runs the action inside the OS sandbox

Audit Recorder logs the outcome

Prepared actions

Built-in tools that cause external side effects do not execute actions immediately. Instead, they produce a PreparedActionDraft structure containing:

  • The requested action name
  • Target resources (file paths, network hosts, environment keys)
  • Intended mutation details (exact line replacements, shell command line)
  • Pre-image hashes of targeted files on disk

The permission engine evaluates this draft against active consent modes, system policies, and sandbox profiles before anything runs.

Built-in tools

Seepient ships with 15 built-in tools organized by capability:

ToolCategoryTypeDescription
read_fileFilesystemRead-onlyReads file contents with line range slicing
write_fileFilesystemEffectfulCreates new files or writes complete file contents
edit_fileFilesystemEffectfulPerforms targeted contiguous edits with line verification
execute_shell_commandSystemEffectfulExecutes shell commands inside the OS sandbox
get_current_datetimeSystemRead-onlyReturns accurate system time and timezone
web_searchNetworkRead-onlyQueries Tavily for real-time web results
read_websiteNetworkRead-onlyExtracts readable markdown from HTTP URLs
take_screenshotBrowserRead-onlyCaptures full-page or element screenshots
send_emailCommunicationEffectfulSends email messages via configured SMTP servers
send_notificationCommunicationEffectfulSends webhook alerts to Feishu, DingTalk, Slack, or Discord
generate_imageMediaEffectfulGenerates images via Fal, Google, or OpenAI backends
render_widgetUIPresentationRenders interactive tables, charts, or forms in the TUI
manage_todosPlanningLocal stateManages the agent's internal task tracking list
optimize_promptAnalysisRead-onlyAnalyzes and refines prompts for upstream models
use_skillOrchestrationContextLoads and activates an external skill folder

TIP

For complete parameter schemas, configuration requirements, and code examples for all 15 tools, see the Built-in Tools Reference.

Read vs effectful tools

  • Read-only tools (such as read_file, web_search, and get_current_datetime) run without prompting in standard consent modes. They cannot alter files or execute arbitrary host instructions.
  • Effectful tools (such as write_file, edit_file, and execute_shell_command) mutate disk state or execute code. In always-ask mode, they prompt for confirmation with a clear visual preview of the command or file diff.

Dry runs

You can run commands with the --dry-run flag. In dry-run mode:

  1. The model plans and reasons normally.
  2. Tools generate prepared action drafts.
  3. The execution boundary logs the action to stdout or the TUI without applying mutations to the filesystem or spawning shell processes.

Next steps

Released under the Business Source License 1.1.