Most agent launches assume a repository. This batch does not: the work is legal research, job applications, bookkeeping, packaging design, language learning and image provenance, and the interesting engineering questions are about trust and containment rather than code generation.
An agent that edits itself
Nodecode is the strangest entry and the most ambitious. It is an agent that rewrites itself, running as one process in SBCL with shells attached. The model has a single tool: it evaluates Lisp in the running program and keeps what it defines. Told it cannot read PDFs and should teach itself, it defines read-pdf, keeps it, and reads the invoice. Setup looks for the other coding agents already on the machine and, with permission, imports their keys, instructions, memory, skills, MCP servers and jobs, so it inherits what you already configured elsewhere.
Work that is not writing code
Junior is an open-source AI junior engineer for senior engineers that runs on your Mac. It watches GitHub, Slack, Linear and your other tools, brings you what needs you, takes the rest, and asks before anything goes out. It runs on Claude Code or Codex and is meant to need you a little less each week.
Seekter is a job-search agent that runs inside Claude Code. It searches job sources every day, filters postings against your own rules for location, visa, sectors, seniority and language, fills forms in your own Chrome, and keeps every application and skip as a markdown file you can read, grep and diff. The rules it was built under are the interesting part: dedup before every form, never guess an answer, never invent an anecdote, never touch a CAPTCHA or a password, and never fill LinkedIn Easy Apply.
Documents, money and terms
CaseCite is a self-hosted legal research platform with eight tools over one corpus: research with semantic retrieval and BM25 rescoring plus inline citations, a contract pipeline that compares two to four contracts clause by clause against a baseline and exports a redline, drafting grounded in documents you pick with a section plan you review first, a case view over CourtListener opinions, and an authority map. Documents, index and chat history stay on your own servers and AI features run under your own keys with no telemetry.
Orchard is cloud storage for a project set up by asking Cursor, with intelligences on top: investment agreement review for US startups covering SAFEs, convertible notes and term sheets with risks and questions for your lawyer, forecasting from a CSV of dates and numbers, image restoration, writing in your style, a Google Meet notetaker and finding tax breaks. It is explicit that its legal output is preparation for counsel and not a legal opinion.
CoreXTrader is a Windows trading terminal with a Strategy IDE, a TDSL agent for AI-assisted coding in its trading DSL, AI market analysis through the Inspector, and a strategy tester with tick-level backtesting, connecting to cTrader Open API brokers with white-label paths for firms.
Making, learning and cleaning
Drint is browser-based packaging design software. You pick a structure, set length, width and height, and the die builds itself; you design on the flat die and the box folds in 3D as you draw, with foil, emboss and kraft rendering like the real stock, and the file goes to press as PDF/X-4 in CMYK with the die on its own layer. It also exposes dielines and 3D mockups as an MCP server.
RainbowSay teaches a language from shows and podcasts. It listens along with the video, pauses at the end of each sentence and shows a card with the translation, a vocabulary list with romanized readings and a grammar note, replays the line with a word-by-word highlight, answers questions about it, then resumes. It listens to audio only and never records your screen.
Free AI Wiper goes the other way and removes traces. It strips supported AI metadata — C2PA manifests, generator records, XMP provenance — from images, documents, video, audio and text in the browser, with visible pixels left unchanged, no upload and no recompression, and it re-inspects the rebuilt file before offering the download.
What the non-code agents have in common
Every entry here has to answer the same two questions before it is useful: what is this agent allowed to touch, and what record does it leave. Nodecode answers by putting everything in one process and asking permission before it imports anything. Junior answers by asking before anything goes out. Seekter answers by writing a markdown file for every application and every skip, then refusing to guess an answer or fill a CAPTCHA. CaseCite answers by keeping the corpus on your own servers. Free AI Wiper answers by never uploading at all.
The pattern is not politeness. An agent that leaves no record cannot be audited, and an agent that cannot be audited is a poor fit for forms, contracts or money. The tools that get this right treat the trace as the product rather than the side effect, which is also why the most convincing ones here are the ones that show you the file they wrote.