Ultimus
A sovereign, local-first agentic coding harness. Ultimus runs the same agent loop against any model, hosted or fully local, so one owner can drive agentic development at depth without provider lock-in or a required cloud dependency. It benchmarks near Claude Code while running on open-source models.
# one owner, one governed loop
$ ultimus "add rate limiting to the API"
read → edit (AST-validated) → bash (sandboxed)
approve? the harness asks before it acts
Running on a local model. Nothing left the machine.
Own the loop, not a wrapper
Most coding agents are thin wrappers around one provider's API. Ultimus is built the other way around. It treats the agent loop as commodity and invests where the durable value is: portability across models, reliable edits, and safe autonomy.
That is what lets a single owner drive real agentic development, at a pace that used to take a team, without handing control or their code to someone else's cloud.
What Ultimus is built for
Sovereignty
One configuration points the same loop at a hosted model, a local model, or any provider. No lock-in, and no cloud dependency you did not choose.
Speed
Intent-based tool filtering and retrieval-pruned context keep each step lean. Minimal dependencies, streaming throughout.
Depth
Edits are validated against the syntax tree, a test-and-verify step runs inside the loop, and tiered memory persists across a session.
What one owner gets
Model sovereignty
Run any model through one gateway: hosted, or local via vLLM, Ollama, or llama.cpp. Switch the whole backend with a single flag, and keep sensitive code on your own hardware.
Safe autonomy
Human-in-the-loop approval on sensitive tools, sandboxed execution, and edits applied against the syntax tree with undo. Automation you can let run.
Grounded, not guessed
Ultimus reads real documentation and pages and cleans them before they enter context, behind an untrusted-data boundary, so a weaker model does not invent an API from a stale memory.
Depth that persists
An append-only session log, tiered memory, context compaction, subagents, and MCP support keep a long task coherent from start to finish.
Attested output
What Ultimus generates can carry a provenance mark and a signed record, so the origin of agent-written code and documents is verifiable rather than assumed. It shows where an output came from; it is not a legal guarantee.
How it is put together
The loop is small and legible on purpose. Each layer does one thing, and every model reaches the loop through a single gateway, so the backend is a configuration choice rather than a rewrite.
CLI / TUI
Agent loop tool registry · permission gate · tool filter · subagents
Session append-only event log · tiered memory · compaction
Model router → one gateway → any model (hosted or local)
Tools read · write · edit (AST) · bash (sandboxed) · search · web fetch
The engine of the loop
Ultimus is the hub. The other three instruments are what it runs on, so the agent is grounded in the outside world, aware of your own code, and governed as it works.
It pulls technical intelligence from Gerolamo, reads your codebase through Prelude, and enforces policy and signs high-consequence actions through Zephyr. What it generates carries verifiable provenance. Agentic depth, grounded and governed, with a clear record.
Put the engine to work.
Ultimus runs inside the full loop we operate for you. If you want agentic development that stays sovereign, governed, and fast, we are glad to walk you through it.