# AI Tools Three modules provide the tools an LLM can call: `mcp_tools` exposes a large built-in tool set over MCP, `rpc_tools` turns `@rpc`-registered app methods into tools on demand, and `ai/subagent` runs isolated subagent sessions with restricted tools. The AI orchestrator merges them into the tool loop.

import "deft/stdext/mcp_tools" as mcp-tools
import "deft/stdext/rpc_tools" as rpc-tools
import "deft/stdext/ai/subagent" as subagent

## MCP Tools `mcp_tools` is the MCP tool server. Tools are plain typed functions with `@doc`; `register-tools` creates the `:deft` MCP server and adds them, and `start-mcp-server` serves `/mcp` over HTTP. | Group | Tools | |---|---| | Help / discovery | `help-lookup`, `list-help-topics`, `list-stdext`, `list-apps` | | Codebase | `read-source`, `grep-codebase`, `search-codebase`, `find-files`, `find-references`, `rename-symbol` | | Editor bridge | `editor-context`, `editor-open-file`, `editor-insert`, `editor-replace-selection`, `editor-replace-buffer`, `editor-goto-line`, `editor-get-selection`, `editor-get-buffer`, `editor-get-buffer-range` | | Git / process | `git-log`, `git-diff`, `run-command`, `run-tests`, `run-diff`, `run-patch`, `write-patch` | | Files | `create-file` | | Metadata | `symbol-lookup`, `fts-search`, `recently-changed` | | Runtime | `eval-dft`, `web-fetch` | | Plumbing | `call-tool`, `get-tools-openai`, `get-tools-anthropic`, `register-tools`, `start-mcp-server` |

[mcp-tools/register-tools]                   # idempotent
set port [mcp-tools/start-mcp-server 7433]   # /mcp endpoint, tries 7433..7441
set r [mcp-tools/call-tool "help-lookup" %{ "topic" "syntax" }]

`call-tool` is the in-TUI dispatcher used by chat; the HTTP endpoint speaks the MCP protocol. The metadata tools open `.deft/meta.db` lazily — see Meta. ## RPC Tools `rpc_tools` is the complementary, opt-in source: it converts methods registered with `@rpc` into provider-format tool schemas and dispatches calls through `[rpc runtime method …]`. Exposure is per app (runtime), never per method. Config shape (delivered by the chat app):

:chat/rpc-tools %{ :include @{ "editor" "diff" } :exclude @{} }

The shipped default include is empty, so nothing is exposed until the user opts in. Tool names are `runtime__method` (collision-free). | Def | Description | |---|---| | `list-methods` | Registered methods filtered by the enabled apps | | `get-tools-openai` / `get-tools-anthropic` | Schemas in provider format | | `describe` | JSON string of the OpenAI-format tools | | `call-tool` | Dispatch a call, coercing string args to declared types | | `config-filter` | Read the `:chat/rpc-tools` config |

[rpc-tools/get-tools-openai %{ :include @{ "editor" } }]
[rpc-tools/call-tool "editor__cmd-open-file" %{ "path" "app.dft" }]

## Subagents `ai/subagent` runs isolated sessions with their own system prompt and a restricted tool allow-list, exposed to the parent model as the `run-subagent` MCP tool. Built-ins: | Name | Focus | Max iterations | |---|---|---| | `code-explorer` | read/search tools | 8 | | `code-reviewer` | review-focused tools | 5 | | `code-writer` | create/edit/test tools | 12 | | `fast-responder` | no tools, fast model | 1 | | Def | Description | |---|---| | `run-named` | Run a built-in subagent; inherits the parent cfg | | `run-adhoc` | Custom system prompt + tool allow-list | | `register-tool` | Idempotently add `run-subagent` to the `:deft` MCP server | | `list-names` / `list-all` / `lookup` | Registry introspection | | `set-provider-cfg` | Stash provider/cfg for the MCP handler (called by chat) | Subagents return `[subagent:<name>] <answer>`. ## See Also - AI — provider registry and orchestration - MCP — the protocol and `mcp/*` stdlib primitives - RPC — `@rpc` method registration - Chat app help — the primary consumer