- llm_core — provider-agnostic LLM orchestration: composable pipelines, TOML configuration, CLI providers, in-process agent loops, structured output.
- llm_toolkit — the base code tools an agent works with (read, write, edit, bash, grep, glob, http) and a composable tool resolver.
- comm_bus — context assembly for LLM conversations: keyword-triggered entries and token budgets.
- agent_engine — a minimal agent runtime with a host-owned turn loop.
- a2a_engine — the Agent2Agent protocol as pure specification: types, JSON-RPC and SSE codecs, a transport behaviour.
- foresight — semantic memory: retain, recall, reflect. A port of Vectorize's Hindsight (MIT). On Hex as
foresight_memory.
- workflow_stem — one runtime for stepwise, state-machine and flow workflows.
- mobus_stepwise — the stepwise and flow engine underneath it.
- bee — dependency-aware work coordination: a work graph, an allocation tree, and queries that say what they left out. On Hex as
bee_tracker.
- attr_engine — schema-driven content modelling over Ecto and JSONB.
- mobus_money — money that carries its currency and an explicit rounding rule.
- mobus_dataview, mobus_scope — directory-view settings; a lazy capability registry for LiveView.
- voice_engine — speech-to-text and text-to-speech behaviours with pluggable adapters.
- greenmail, fosferon_telemetry — a local mail server for tests; shared failure-telemetry conventions.
All are on Hex under the MIT licence.
- elixir-context — project-wide search and call-graph context for Elixir, served to coding CLIs over MCP. MIT.
For more information: fosferon.com.