All blogs

What Shipped in Zorp 0.4

Until August, Zorp was a terminal. You wrote a pre-registration file, ran zorp validate, and read a ledger. That is still the core of it, and the part we care most about being right. But a research tool nobody can open is a research tool nobody uses, and the pre-registration argument does not travel well through a --help screen.

The 0.4 line, v0.4.0 on 25 August and v0.4.1 on 1 September, gave Zorp a browser. Here is what is in it, what is deliberately still gated, and what we have not measured.

Forty seconds of it

Onboarding

Installing zorp, starting zorp-web, picking a model, and running the first conversation through to an artifact. Silent, unedited, single take.Source: zorp v0.4.1 on macOS, recorded 1 September 2026, brightness lifted for legibility

The chat UI

zorp-web serves a local chat interface on 127.0.0.1:7777. It is a real client, not a demo page: answers stream as the model writes them, markdown renders, the send button becomes a stop button and the stop actually reaches the model rather than just hiding the output.

Anything a run writes opens in an artifact pane beside the conversation. The pane reads PDFs and office formats rather than framing them in a viewer you cannot select text out of, and it resizes and hides. In 0.4.1 the workspace became two draggable halves, with Files as a picker rather than a list.

Conversations are first-class. Each gets a name in its own column, lives in the address bar so a link reopens it, and, as of 0.4.1, can be deleted from the sidebar. Messages queue while a turn is still running instead of being dropped.

Model choice is in the UI. A local Ollama, a frontier API, or the oMLX preset added in 0.4.1 with API key support. A finished answer copies out in one click, including formats aimed at Claude, Codex, or Gemini, because handing a result to another agent is what people actually do with it.

Search, voice, and recall

Native web search landed with Tavily as the first provider. The UI says whether the agent can search and means it, rather than leaving you to guess whether an answer came from evidence or from the weights.

Voice input runs Qwen3-ASR over loopback, sets itself up on a click, and previews the transcription live while you are still talking. No audio leaves the machine.

Recall indexes your conversations automatically and searches them on this machine only. Every conversation feeds a memory that is quoted, never summarized, so what comes back is what was said and not a model’s compression of it.

The parts that check the work

An adversarial review panel runs several agents against a draft at once. Self-critique audits a draft against the evidence record before delivery, and 0.4.1 continued that work and let the safety reviewer auto-approve where nothing is at stake.

Underneath, the discovery layer, Aryabhatta, was built out over seven steps in 0.4.0: a re-run gate, an anomaly ledger, anomaly families, a computed go/no-go, and a calibration run over a real corpus. 0.4.1 added hypothesis search on an ERBGA substrate, gated behind a flag, and fixed investigate so a run’s max_steps is recorded and the real ledger can reach three condition keys.

Zorp also reads Claude Code skills, and can grant them nothing at all.

Install, with no toolchain

The 0.3.x line already dropped the Rust requirement. 0.4 ships prebuilt zorp, zorp-agent and zorp-web binaries plus the static web UI, so neither Rust nor Node is needed.

curl -fsSL https://raw.githubusercontent.com/aviskaar/zorp/main/install.sh | bash

Or take the archive for your platform with its .sha256, put both in one directory, and check it before you trust it:

shasum -a 256 -c zorp-*.tar.gz.sha256    # sha256sum -c on Linux

Then:

zorp-web    # http://127.0.0.1:7777

What is not in this

The prebuilt binaries do not include the research capabilities. validate, investigate, co-write and deliver sit behind the research feature and still require building from source. Hypothesis search is gated. Neither of those is an oversight; they are the parts we are least willing to ship to someone who has not opted into them.

We also have no new numbers. 0.4 is a surface release: it makes the existing machinery reachable and does not claim to make it more correct. The groundedness evaluation, the citation-accuracy figure, the hallucination rate, and the comparison against a plain language model are all still missing, and they are still the numbers that would actually settle the question.

Zorp is pre-alpha. Interfaces change. The changelog is the honest version of this post.