Distribute and run LLMs with a single file.
C++ cross-platformggufllama-cpplocal-ailocal-inferencelocal-llmopen-source-aisingle-file-executablespeech-to-text
#-*-mode:makefile-gmake;indent-tabs-mode:t;tab-width:8;coding:utf-8-*-┐
#── vi: set noet ft=make ts=8 sw=8 fenc=utf-8 :vi ────────────────────┘
#
# Copyright 2026 Mozilla.ai
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# BUILD.mk for agentfile (agentic CLI binary)
#
# agentfile is a cosmocc-compiled APE binary that runs an agentic loop
# end-to-end from a single CLI invocation. It builds on pieces the tree
# already compiles: agent.cpp's Agent/Model (loop + llama.cpp wrapper),
# llama.cpp's server-tools objects (the tool implementations), llama.cpp.a,
# and the TOOL_LLAMAFILE_OBJS providing GPU dispatch + zip filesystem.
#
PKGS += AGENTFILE
# ==============================================================================
# Source files
# ==============================================================================
AGENTFILE_SRCS_CPP := \
agentfile/agentfile.cpp
# _DEP_HDRS, not _HDRS: deps.mk collects $(PKG)_SRCS/_HDRS/_INCS for
# mkdeps, which would scan these headers and warn about includes it cannot
# resolve (agent.cpp and llama.cpp declare no header lists). Like
# AGENT_CPP_DEP_HDRS, this list only triggers recompiles.
AGENTFILE_DEP_HDRS := $(wildcard agentfile/*.h) \
$(wildcard agentfile/callbacks/*.h) \
$(wildcard agentfile/tools/*.h)
# Headers from other projects that agentfile.cpp includes (agent.cpp's
# list plus server, mtmd, vendored and llamafile ones). A submodule bump
# changes them without touching agentfile sources, and a stale agentfile.o
# then fails to link against the new objects.
AGENTFILE_EXT_HDRS := \
$(AGENT_CPP_DEP_HDRS) \
$(wildcard llama.cpp/tools/server/server-*.h) \
$(wildcard llama.cpp/tools/mtmd/*.h) \
llama.cpp/vendor/cpp-httplib/httplib.h \
llama.cpp/vendor/sheredom/subprocess.h \
llamafile/llamafile.h
# llama.cpp server-tools objects the adapter needs (tool implementations
# plus the TUs providing their symbols; no server.cpp — that has main()).
AGENTFILE_SERVER_OBJS := \
o/$(MODE)/llama.cpp/tools/server/server-tools.cpp.o \
o/$(MODE)/llama.cpp/tools/server/server-common.cpp.o \
o/$(MODE)/llama.cpp/tools/server/server-queue.cpp.o \
o/$(MODE)/llama.cpp/tools/server/server-mcp.cpp.o \
o/$(MODE)/llama.cpp/tools/server/server-http.cpp.o
AGENTFILE_OBJS := $(AGENTFILE_SRCS_CPP:%.cpp=o/$(MODE)/%.o)
# ==============================================================================
# Include paths
# ==============================================================================
AGENTFILE_INCLUDES := \
-iquote agentfile \
-iquote llamafile \
-iquote agent.cpp/src \
-iquote llama.cpp/common \
-iquote llama.cpp/include \
-iquote llama.cpp/ggml/include \
-iquote llama.cpp/tools/server \
-iquote llama.cpp/tools/mtmd \
-isystem llama.cpp/vendor
# ==============================================================================
# Compiler flags
# ==============================================================================
# COSMOCC enables the cosmo_args("/zip/.args") call in agentfile.cpp, as in
# llamafile/BUILD.mk; cosmocc itself only predefines __COSMOCC__.
AGENTFILE_CPPFLAGS := $(AGENTFILE_INCLUDES) \
-DLLAMAFILE_VERSION_STRING=\"$(LLAMAFILE_VERSION_STRING)\" \
-DCOSMOCC=1
# cpp-httplib is built with its Mbed TLS backend (see the HTTPS section in
# llama.cpp.patches/llamafile-files/BUILD.mk). The macro changes httplib
# class layouts, so agentfile objects that include httplib.h (http_fetch,
# web_search) must define it too, and the executable must link mbedtls.a.
AGENTFILE_CPPFLAGS += \
-DCPPHTTPLIB_MBEDTLS_SUPPORT \
-isystem third_party/mbedtls/include
# common/subproc.h (reached through server-common.h) switches subprocess_s
# between its real and dummy definition on LLAMA_SUBPROCESS. The server and
# common objects linked below are built with it (see the subprocess section
# in llama.cpp.patches/llamafile-files/BUILD.mk), so agentfile must agree.
AGENTFILE_CPPFLAGS += -DLLAMA_SUBPROCESS
o/$(MODE)/agentfile/%.o: agentfile/%.cpp agentfile/BUILD.mk $(AGENTFILE_DEP_HDRS) $(AGENTFILE_EXT_HDRS)
@mkdir -p $(@D)
$(COMPILE.cc) $(AGENTFILE_CPPFLAGS) -o $@ $<
# ==============================================================================
# Executable
# ==============================================================================
# Links:
# - agentfile sources (agentfile.cpp)
# - agent.cpp static lib (Agent, Model, ModelWeights)
# - llama.cpp static archive (ggml, llama, common)
# - llamafile objects for runtime GPU dispatch + zip filesystem
# - LLAMAFILE_METAL_SOURCES: embedded Metal kernel sources (macOS)
# - TINYBLAS_CPU_OBJS for matmul kernels
# - HTTPLIB_OBJS: cpp-httplib implementation (server tools, http_fetch,
# web_search)
o/$(MODE)/agentfile/agentfile: \
$(AGENTFILE_OBJS) \
o/$(MODE)/agent.cpp/agent.cpp.a \
o/$(MODE)/llama.cpp/llama.cpp.a \
o/$(MODE)/third_party/mbedtls/mbedtls.a \
$(AGENTFILE_SERVER_OBJS) \
$(TOOL_LLAMAFILE_OBJS) \
$(LLAMAFILE_METAL_SOURCES) \
$(TINYBLAS_CPU_OBJS) \
$(HTTPLIB_OBJS)
@mkdir -p $(@D)
$(LINK.o) $(AGENTFILE_OBJS) o/$(MODE)/agent.cpp/agent.cpp.a $(AGENTFILE_SERVER_OBJS) $(TOOL_LLAMAFILE_OBJS) $(LLAMAFILE_METAL_SOURCES) $(TINYBLAS_CPU_OBJS) $(HTTPLIB_OBJS) o/$(MODE)/llama.cpp/llama.cpp.a o/$(MODE)/third_party/mbedtls/mbedtls.a $(LOADLIBES) $(LDLIBS) -fopenmp -lpthread -o $@
# ==============================================================================
# Main target
# ==============================================================================
.PHONY: o/$(MODE)/agentfile
o/$(MODE)/agentfile: o/$(MODE)/agentfile/agentfile