Tools you probably do not know yet. No ChatGPT, no Canva: you already have those. For each one: what it does, what it does not do, and the link to its repo or vendor. Type what you are trying to do, the list narrows.
Nothing for that search. Tell us which tool is missing.
Turns a PDF into clean text, tables and equations included, without losing structure. Runs on your machine.
What it does not do Does not read handwriting, and does not straighten a crooked scan. Without a GPU, no fast processing.
github.com/datalab-to/marker →Same job, from IBM: PDF, Word, PowerPoint and images into a format models read well.
What it does not do A developer library, no interface: you call it from a program.
github.com/docling-project/docling →Turns a whole site into clean text for a model, following links and handling JavaScript.
What it does not do Does not tell you whether the target site allows scraping: that check is not in the tool.
www.firecrawl.dev →Around fifty PDF operations, merge, split, sign, OCR, on your own machine, with nothing uploaded anywhere.
What it does not do No AI at all: no content reading, no summary, no data extraction. File handling only.
github.com/Stirling-Tools/Stirling-PDF →A coding agent in the terminal that works inside your git repo and commits every change.
What it does not do No graphical interface, and no model included: it calls the API you choose, at your cost.
aider.chat →The open-source coding assistant for VS Code or JetBrains, with the model you choose, local included.
What it does not do Ships no model: you pick and wire one at install time.
www.continue.dev →An agent that opens a terminal and a browser to finish a development task on its own.
What it does not do Really runs the commands on the system it is given: it does not simulate them.
github.com/All-Hands-AI/OpenHands →One way to call a hundred model providers, with automatic failover when one goes down.
What it does not do Does not improve answers: it routes, it does not change what the model produces.
github.com/BerriAI/litellm →Forces a model to answer in a structure you describe, and retries until the answer fits.
What it does not do It guarantees the shape, never the substance: a valid structure can hold a wrong number.
github.com/567-labs/instructor →See what each step of an agent actually did, what it cost and where it failed. Self-hostable.
What it does not do It shows the failure, it does not repair it. And instrumenting means touching the code.
langfuse.com →Chain tools and models into visual automations, on your own server if you want.
What it does not do Ships no business logic: the chaining and the failure handling are built node by node.
n8n.io →Memory for agents that tracks where each fact came from, with an MCP server ready to plug in.
What it does not do One node only, and memory is filled by hand: nothing is extracted from conversations on its own.
github.com/JordyZomer/lemmalog →One command to download and run a model on your computer. Nothing leaves the machine.
What it does not do Does not go beyond the machine's memory: a model larger than the available memory does not load.
ollama.com →The same idea with a window: find a model, load it, chat, no command line.
What it does not do Closed source, and no model of its own: what comes out comes from the model you load.
lmstudio.ai →The engine that serves an open model to many users at once without melting the GPU.
What it does not do Does not target a laptop: it is built for a GPU server.
github.com/vllm-project/vllm →Guided assumptions in, formatted financial forecast out, in the shapes a bank or investor expects.
Price from $7 a month, refundable within fifteen days
What it does not do It formats, it checks neither your assumptions nor your figures.
upmetrics.co →Describe a home in one sentence, get a floor plan, refine it in conversation, see it in 3D.
Price free to start
What it does not do Does not produce a regulatory document: no stamped plan, no permit file.
www.maket.ai →Interior floor plan in 2D then a 3D render, aimed at layout and furnishing.
What it does not do Interior layout: no regulatory document, no structural calculation.
planner5d.com →A deck written in markdown, versioned in git, with code that runs inside the slides.
What it does not do No visual editor: slides are written in markdown.
sli.dev →Same idea, simpler, and it exports to PDF or PowerPoint from a markdown file.
What it does not do No free placement of elements: you follow the theme, or you write CSS.
marp.app →Build your own image pipeline by wiring boxes, with control the hosted services do not offer.
What it does not do No 'one sentence, one image' box: everything goes through a node graph, and it needs a GPU.
github.com/comfyanonymous/ComfyUI →Image generation inside a real painting program: you paint, select, regenerate the area.
What it does not do Does not run on its own: it needs Krita, and the model runs on your machine.
github.com/Acly/krita-ai-diffusion →Local transcription with word-level timestamps and speaker separation: who said what, and when.
What it does not do No interface, everything goes through the command line. And no live transcription: it works on an already recorded file.
github.com/m-bain/whisperX →The same transcription, in C, light enough to run on a laptop or a phone.
What it does not do No interface, and no speaker separation: it gives you text, not who is speaking.
github.com/ggml-org/whisper.cpp →A synthetic voice that runs on your own hardware, Raspberry Pi included, in many languages.
What it does not do Does not clone a voice and takes no emotion instruction: the voice is the one in the model you pick.
github.com/OHF-Voice/piper1-gpl →Ask your database a question in plain words: it writes the SQL, runs it and shows the query.
What it does not do Does not know your database out of the box: you supply the schema and the examples.
vanna.ai →Query a spreadsheet or a database in plain language, and get the chart with the code that made it.
What it does not do Runs the code it writes, and does not sandbox that execution itself.
github.com/sinaptik-ai/pandas-ai →A conversational search engine you host yourself, which cites its sources and sends nothing to a third party.
What it does not do No hosted service to open: you install it, and wire a search engine behind it.
github.com/ItzCrazyKns/Perplexica →One address to try dozens of models, including the ones released early under a pseudonym.
What it does not do Hosts no model: it routes to the providers, and follows their prices and their outages.
openrouter.ai →'What it does not do' is not a criticism: it is the tool's boundary, written as a plain fact, so you know before clicking whether it fits. The directory does not rank and does not score, and it deliberately leaves out the tools everyone already has. A missing tool is not a bad tool: it has not been checked yet. Prices move, they are given as the vendor displayed them on the day of the check.
Geneva, Switzerland. Write to hello@newsworldai.xyz.