Skip to content
Tandem Engine

Tools Reference

The Tandem Engine tool registry currently exposes the following tools.

File Operations

  • read: Read file contents.
    • Input: path (string)
  • write: Write file contents (overwrites).
    • Input: path (string), content (string)
  • edit: String replacement in a file.
    • Input: path (string), old (string), new (string)
  • glob: Find files by pattern.
    • Input: pattern (string, e.g., src/**/*.rs)

For coding tasks, agents usually read first, use edit for narrow replacements, write for new or rewritten files, and apply_patch when the change is easier to review as a unified diff.

  • grep: Regex search in files.
    • Input: pattern (string), path (string, root directory)
  • websearch: Search the web (powered by Exa.ai).
    • Input: query (string), limit (integer)
  • codesearch: Semantic code search (if configured).
  • memory_list: List stored memory records/chunks from the current session, project, or global scope.
    • Input: optional scope controls such as tier, session_id, project_id, limit, allow_global
  • memory_search: Search memory by query across the current session, project, or global scope.
    • Input: query plus optional scope controls such as tier, session_id, project_id, limit, allow_global
  • memory_store: Persist content into session, project, or global memory.
    • Input: content plus optional scope and metadata such as tier, session_id, project_id, source, metadata, allow_global
  • memory_delete: Delete a stored memory record/chunk within the current allowed scope.
    • Input: chunk_id (or id) plus optional scope controls such as tier, session_id, project_id, allow_global

Web

  • webfetch: Fetch URL and return structured Markdown/JSON output.
    • Input: url (string), optional mode, return, max_bytes, timeout_ms, max_redirects
  • webfetch_html: Fetch URL and return raw HTML text.
    • Input: url (string), optional max_bytes, timeout_ms, max_redirects

System

  • bash: Run shell commands (PowerShell on Windows, Bash on Linux/Mac).
    • Input: command (string)
  • mcp_list: List the configured and connected MCP servers and their discovered tools.
    • Input: none
    • Use this first when an agent needs to discover MCP access.
    • It returns a structured inventory in one result, which keeps the agent from carrying a full MCP tool dump in context unless it actually needs one.
    • If the needed server or tool is missing, stop and tell the user to add or connect the MCP instead of inventing a capability.
    • If a workflow request depends on an unavailable MCP, the right response is to ask for that MCP, not to silently switch to a different tool.
  • mcp_list_catalog: List the embedded MCP catalog with connection overlay for gap analysis.
    • Input: none
    • Use this after mcp_list when you need to distinguish connected, cataloged, disabled, and uncataloged servers.
  • mcp_request_capability: Request human approval for an MCP capability gap without attempting to connect or execute it directly.
    • Input: agent_id (string), mcp_name (string), optional catalog_slug (string), rationale (string), optional requested_tools (array of strings), optional context (object), optional expires_at_ms (integer)
    • Use this to file an approval request when the required MCP is not connected or is uncataloged.
    • Edition note: this keeps the same tool name in OSS builds, but can return an explicit premium-feature error when managed governance is unavailable.
  • mcp_debug: Call an MCP tool directly by URL.
    • Input: url (string), tool (string), optional args (object), headers (object), timeout_ms (integer), max_bytes (integer)
    • Use this when you already know the MCP server URL and the tool name you want to invoke.
    • This tool does not discover tools or search the registry; it only calls a named MCP tool.
  • todo_write: Update the Todo/task list.
    • Aliases: todowrite, update_todo_list
  • task: Update the current task status.
  • question: Ask a structured question to the user and wait for input.
  • spawn_agent: Spawn an agent-team worker instance (runtime/policy gated).
    • Input: mission/spawn payload (e.g., missionID, role, templateID, source)
  • teamcreate: Create/register an agent-team context for coordinated teammate tasks.
    • Input: team metadata (e.g., team_name, description, agent_type)
  • taskcreate: Create teammate task records in a team context.
    • Input: task payload (e.g., team_name, name, description)
  • taskupdate: Update teammate task status/notes/progress in a team context.
    • Input: task update payload (e.g., team_name, task_id, status, notes)
  • tasklist: List tasks for a team context.
    • Input: optional filters (e.g., team_name, status filter)
  • sendmessage: Send mailbox-style message/task prompt to one or more teammates.
    • Input: message payload (e.g., team_name, to, content, summary)

MCP Discovery Endpoints

These are HTTP endpoints, not engine tools:

  • mcp_list - Engine tool that returns the current MCP inventory snapshot.
  • GET /mcp/tools - List discovered tools from connected MCP servers.
  • GET /tool/ids - List all engine tool IDs, including built-ins and MCP tools.

If you need to find a tool by keyword, fetch one of those lists and filter locally. Tandem does not provide a separate public “search the installed registry” endpoint, but mcp_list does give the agent a structured inventory of connected servers and tools, which is the lower-context first step before selecting a specific MCP tool.

Specialized

  • skill: Execute a skill.
  • apply_patch: Apply a unified diff patch.
  • batch: Execute multiple tools in a batch.
  • lsp: Interact with the Language Server Protocol.