Tools

Built-in tools

A fresh prokop init can work with files, search code, run commands, fetch the web, ask structured questions, and keep a task list without installing a separate tool pack. These are the tools of the Prokop runtime; Claude Code and Codex use their own, shown with the same names and views in the workspace. Delegation, messaging, and group tools come with the workspace tools and are the same on every runtime; see Built-in MCP entries.

Files and search

Tool	Purpose
read-file	Read text files with line numbers or list a directory
file-to-markdown	Convert PDF, Office, LibreOffice, and ZIP files into readable Markdown
write-file	Create a file or deliberately replace its full contents
edit	Make one or more targeted replacements in a file, with revision checking
glob	Find files by path pattern
grep	Search file contents with regular expressions

Editing tools respect workspace boundaries and the permission mode. Existing files are normally changed with targeted edits instead of full rewrites.

Shell

The shell tool runs each command in its own process with your shell (zsh, bash, dash, ksh, or sh), so commands started side by side don’t wait on each other. A command still running after the wait, 60 seconds unless the agent asks for more, keeps running in the background. The agent gets the output so far and a message in the conversation when the command finishes, with the output attached as a terminal chip. It can read more output or stop the command, and the full output is kept in a file it can search.

Background commands stop when the session is archived or deleted, when you undo or edit a message, and when the server stops. Interactive prompts are not supported; agents use non-interactive flags instead.

Commands ask for approval only when there is a concrete concern, such as a path outside the workspace, secrets, or a destructive operation. The approval card highlights the part of the command that needs review and says why. See Permissions.

Web and browser

Tool	Purpose
webfetch	Fetch a URL and convert it to readable text or Markdown
Browser tools	Read tabs and interact with pages through Prokop Browser

For web search, connect a search service as an MCP server: in Settings → MCP Servers for every workspace, or in a workspace’s MCP settings for that workspace only.

Browser tools are built in but not enabled in the bundled agents. Add the ones you need to an agent: browser_navigate, browser_read_active_tab, browser_tab_manage, browser_discover_elements, browser_dom_action, and browser_screenshot. They require the Prokop Browser extension connected to the same server.

Interaction and task management

The question tool renders typed forms for confirmation, free text, single selection, or multiple selection. The todo tool keeps a visible task list for longer work.

Delegation is not a runtime tool anymore: delegate_task and its companions are built-in MCP tools, the same on Prokop, Claude Code, and Codex. See Prokop Delegation.

Memory, skill, and search tools

Tool	Available when	Purpose
memory	Memory is on for the workspace	Manage the workspace’s preferences and project facts
skill_manage	Skill management is on for the workspace	Create, patch, update, or delete workspace skills
agent_memory	The agent’s memory capability is on (default)	Manage the agent’s own memory, which follows it across workspaces
agent_skill_manage	The agent’s skills capability is on (default)	Manage the agent’s own skills
skill	Skills are available	Load one skill’s full procedure
session_search	Always	Find and read earlier conversations in the session, the workspace, or across the agent’s work

Claude Code and Codex sessions get the same memory, session search, and agent skill tools. See Workspaces & sessions for storage and defaults.

MCP tools

Connect an MCP server and its tools appear alongside the built-in ones. Prokop supports local stdio processes and remote Streamable HTTP/SSE servers with optional request headers, and handles OAuth on the server.

Workspace MCP definitions live in <workspace>/.prokopai/mcp.json. See Configuration.

Tools on demand

A Prokop session in a workspace with more than ten MCP tools does not get every schema on every step. The agent sees one tool_search tool with the deferred tools listed by name and server, loads the ones it needs by name or by keyword, and those stay callable for the rest of the session. The built-in entries below are never deferred. Claude Code and Codex sessions already load MCP tools this way on their own.

Built-in MCP entries

These entries need no setup and work the same on Prokop, Claude Code, and Codex. They are not MCP servers to configure: Visuals is switched in MCP settings, the rest follow the agent’s own settings.

Prokop Visuals

One tool, render_html. Agents use it to show a chart, table, diagram, or mockup as a page in their reply, up to 200 KB. The page runs sandboxed, with no access to your login or stored data, and follows your light or dark theme. Subagents and learning reviews don’t get it. Turn it off, or turn off render_html, in MCP settings for every workspace or for one.

Prokop Delegation

An agent has these when it may spawn subagents, whatever its mode.

Tool	Purpose
delegation_capabilities	The agent’s subagent roles with their defaults, the runtimes enabled on this machine with a short model list each, and the limits
delegate_task	Hand a task to a role, on its own runtime and model or on a target the agent picks; isolation: "worktree" starts a top-level session on a new branch in a new managed worktree, isolation: "checkout" one in the current checkout
task_status	Check on a task the agent started
task_cancel	Stop one

A delegated task returns at once by default and its result arrives in the parent session as a new message when it finishes; mode: "wait" blocks for it. A child’s question or approval shows on the session you are looking at. Each agent’s Delegation targets and New sessions settings say which runtime or model choices and which worktree sessions need your yes; see Delegation.

Prokop Messaging

One tool, message_agent: send another of your agents a request, a question, a status, a result, or an FYI. The recipient works on it as itself, in a conversation session in its own home. An agent accepts messages only when its Messaging settings say so, and only agents with a home can; see Agents.

Prokop Groups

Members of a group get group_history, which reads earlier messages of the group beyond what a wake carried, and handoff_to_agent, which passes a distinct next stage to another member and wakes it.

Permissions and asks

Every session, workspace, and scheduled job runs in one of three permission modes:

Mode	Behavior
Standard	Asks when a command or file operation reaches outside the workspace, touches secrets, or is destructive
Extended	Also allows paths outside the workspace; secrets and destructive operations still ask
Full	Runs without asking, except for catastrophic commands

Commands classified as catastrophic, such as rm -rf /*, ask in every mode. All three runtimes follow the same rules. An ask shows the concern and its evidence; for commands, the card shows each pipeline stage, highlights the part to review, and dims the rest. Allow Once is the main action. Grants are scoped and can be reviewed or revoked in the client. In Prokop runtime sessions, allowing for the session or workspace covers the exact command in the same folder, or the exact file, and commands that can only be allowed once can’t be remembered.

Tools can also pause to ask for input, such as a question form. This permission layer is separate from server authentication: authentication controls who can connect, and permissions control what an agent may do after connecting.