ES version is available. Content is displayed in original English for accuracy.
MCP tools don't really work for financial data at scale. One tool call for five years of daily prices dumps tens of thousands of tokens into the context window. And data vendors pack dozens of tools into a single MCP server, schemas alone can eat 50k+ tokens before the agent does anything useful. So we auto-generate typed Python modules from the MCP schemas at workspace init and upload them into the sandbox. The agent just imports them like a normal library. Only a one-line summary per server stays in the prompt. We have around 80 tools across our servers and the prompt cost is the same whether a server has 3 tools or 30. This part isn't finance-specific, it works with any MCP server.
The other big thing was making research actually persist across sessions. Most agents treat a single deliverable (a PDF, a spreadsheet) as the end goal. In investing that's day one. You update the model when earnings drop, re-run comps when a competitor reports, keep layering new analysis on old. But try doing that across agent sessions, files don't carry over, you re-paste context every time. So we built everything around workspaces. Each one maps to a persistent sandbox, one per research goal. The agent maintains its own memory file with findings and a file index that gets re-read before every LLM call. Come back a week later, start a new thread, it picks up where it left off.
We also wanted the agent to have real domain context the way Claude Code has codebase context. Portfolio, watchlist, risk tolerance, financial data sources, all injected into every call. Existing AI investing platforms have some of that but nothing close to what a proper agent harness can do. We wanted both and couldn't find it, so we built it and open-sourced the whole thing.
Discussion (49 Comments)Read Original on HackerNews
I maintain an OSS SDK for Databento market data. A year ago, I naively wrapped the API and certainly felt this pain. Having an API call drop a firehose of structured data into the context window was not very helpful. The tool there was get_range and the data was lost to the context.
Recently I updated the MCP server [1] to download the Databento market data into Parquet files onto the local filesystem and track those with DuckDB. So the MCP tool calls are fetch_range to fill the cache along with list_cache and query_cache to run SQL queries on it.
I haven't promoted it at all, but it would probably pair well with a platform like this. I'd be interested in how people might use this and I'm trying to understand how this approach might generally work with LLMs and DuckLake.
[1] https://github.com/NimbleMarkets/dbn-go/blob/main/cmd/dbn-go...
Most people, before being confronted to it, have no idea how big market data feeds really are: I certainly had no idea what I was getting into. There's a reason all these subscriptions are that pricey.
Here's an example of the pricing for the OPRA feed for Databento you mentioned:
https://databento.com/pricing#opra
We're talking about feeds that sustain 25+ Gb/s and can have spikes at twice or even three times that. And that's only for options market data.
I mean: even people with 25 GB/s fiber (which we can all agree ain't the most common and that's an understatement) at home still can't dream of getting the entire feed.
Having a bandwith big enough, storing, analyzing such amount of data: everything becomes problematic as such scales.
As to me I'm reusing my brokers' feeds (as I already pay for them): it's not a panacea but I get the info I need (plus balances/orders/etc. tied to my accounts).
I’m just noting for interest that firms are applying transformers and other networks at this streaming microstructure level, but specially trained for feature extraction. HRT + Nvidia have some nice videos about it
I will also note that it is insane how much better all the LLMs are at calling MCP tools after just a year, especially the local ones.
One of the reasons I like DuckDB is the scale flexibility. I started with grabbing data and playing on my laptop, then I jumped to a server with high cores and a NAS.
This is a problem with pretty much everything beyond easy single-shot tasks. Even day-to-day stuff, like e.g. I was researching a new laptop to buy for my wife, and am now enlisting AI to help pick a good car. In both cases I run into a mismatch with what the non-coding AI tools offer, vs. what is needed:
I need a persistent Excel sheet to evolve over multiple session of gathering data, cross-referencing with current needs, and updating as decisions are made, and as our own needs get better understood.
All AI tools want to do single session with a deliverable at the end, that they they cannot read, or if they can read it, they cannot work on it, at best they can write a new version from scratch.
I think this may be a symptom of the mobile apps thinking that infects the industry: the best non-coding AI tools offered to people all behave like regular apps, thinking in sessions, prescribing a single workflow, and desperately preventing any form of user-controlled interoperability.
I miss when software philosophy put files ahead of apps, when applications were tools to work on documents, not a tools that contain documents.
Most non-coding AI tools are meant for general consumers who normally don't care if they have to do a new search each session + the hacky memory features try to tackle this over the long term. Also you can always supply it with the updated file at each prompt and ask it to return by updating the file. (if you really want to do with something like ChatGPT).
And I think its a bit hyperbole to extrapolate this to "software philosophy is changing". Like most apps still work on documents/data? Not sure what you meant there
They do care. The vendors don't. Or rather, they're not prioritizing it.
Technically, two of the major players own office suite software, into which they integrate AI, and it's kinda starting to become usable for something. But it is still a bit ridiculous that there's nothing in mainstream tools themselves between single-shot document output and full Computer Use in Claude Desktop. Multi-document edit wouldn't be that hard to make as an extension of "canvas" mode in these tools.
> Like most apps still work on documents/data? Not sure what you meant there
I mentioned mobile. Most apps don't work on data documents, they work in private databases, and collaborate with the OS to keep you from accessing your data directly.
It is absolutely a philosophical change. The core unit of computer use in desktop era was a file. Data belonged in files, which were owned by users. Applications were used to operate on those files. In the mobile era, the core unit of computing is an app, which owns the data, and may or may not graciously let other apps access a partial, non-canonical copy by means of "Share" menu.
This one-shot content generated is quality-less, no-focus. I would not, but cannot help to use the word, garbage. That's why some time you see a picture or a piece of video, or even some content, you can smell it is AI. And AI smell stinks.
The real useful AI content(for this case, the financial goal) generation is a long term efforts, and current AI agent is not good at it. That's why the author created this platform(or App) to help people achieve there.
But most people who one-shot content barely bother to read it, let alone edit or polish. And that's why it stinks. I doubt that better AI agents will solve this. Writing powerful content is actually a pretty hard and subjective task; writing boilerplate marketing copy is very easy and mostly objective.
If the idea is that people who don't understand content will use it to generate content, then its logical to conclude they don't understand why their content stinks as well and will continue to publish the said content.
I feel almost 100% of content published (even on high-tier publications like The Guardian) have AI involved in creating them but their strong editorial processes prevent shit like "It's not LIKE THIS BUT LIKE THIS!" type text to be present 100 times in a 500 word article.
People need to understand writing is a skill and an art. AI can emulate it but so far, no model seem to have genuine creativity or skill in their writing unless you wrestle with it (but at that point it is YOUR skill that you're using to make the AI better).
It's a full stack open-source agent harness (Apache 2.0). Persistent sandboxed workspaces, code execution against financial data, and a complete UI with TradingView charts, live market data, and agent management. Works with any LLM provider, React 19 + FastAPI + Postgres + Redis.
MCP tools don't really work for financial data at scale. One tool call for five years of daily prices dumps tens of thousands of tokens into the context window. And data vendors pack dozens of tools into a single MCP server, schemas alone can eat 50k+ tokens before the agent does anything useful. So we auto-generate typed Python modules from the MCP schemas at workspace init and upload them into the sandbox. The agent just imports them like a normal library. Only a one-line summary per server stays in the prompt. We have around 80 tools across our servers and the prompt cost is the same whether a server has 3 tools or 30. This part isn't finance-specific, it works with any MCP server.
The other big thing was making research actually persist across sessions. Most agents treat a single deliverable (a PDF, a spreadsheet) as the end goal. In investing that's day one. You update the model when earnings drop, re-run comps when a competitor reports, keep layering new analysis on old. But try doing that across agent sessions, files don't carry over, you re-paste context every time. So we built everything around workspaces. Each one maps to a persistent sandbox, one per research goal. The agent maintains its own memory file with findings and a file index that gets re-read before every LLM call. Come back a week later, start a new thread, it picks up where it left off.
We also wanted the agent to have real domain context the way Claude Code has codebase context. Portfolio, watchlist, risk tolerance, financial data sources, all injected into every call. Existing AI investing platforms have some of that but nothing close to what a proper agent harness can do. We wanted both and couldn't find it, so we built it and open-sourced the whole thing.
It saves the payloads into SQLite, maps them, and exposes tools for the model to run python against them. Works very well.
Their charts and UI look pretty, too.
We currently persist logs for every single line of code agent executed in the sandbox. We also have agent trajectory persisted with infra from langchain ecosystem.
That said, i share the same believe that lots of work need to be done for compliance if deploying this for any firm on the street.
>Curious whether LangAlpha has thought about signed execution logs per session.
This is defiantly what we have on the roadmap but feels a little heavy to introduce at early stage.
What I missed from the writeup were some specific cases and how did you test that all this orchestration delivers worthwhile data (actionable and full/correct).
E.g. you have a screenshot of the AI supply chain - more of these would be useful, and also some info about how you tested that this supply chain agrees with reality.
Unless the goal of the project was to just play with agent architecture - then congrats :)
For demo purpose and to attract attention, i was primarily picking some cases with cool visuals (like the screenshot of the AI supply chain you mentioned). we have some internal eval and will try to add more cases in the public repo for reference.
>And don't give me that "hacker culture" bullshit
I couldn’t help but be genuinely curious: if you believe AI is a bubble and aren’t a fan of hacker culture, then why are you here on Hacker News?
great to hear your input anyway!
You might be interesting in what I've been working on. I've discovered giving an autoresearch approach to letting Claude write Claude, it will find lots of alpha everywhere beating SPY buy and hold. It will even find alpha filling in gaps with trading gold ETFs as a hedge. [0] What it really is a bug squashing agent. LLMs will lie and cheat at every move and can't be trusted. 75% (3/4 of agents and code is dedicated to this) of creating agents and using LLMs with financial data is hunting and squashing the bugs and lies.
[0] https://github.com/adam-s/alphadidactic
Is this kind of like a Karpathy 2nd brain for investing then?
As for now, it would be more like how swe working on a codebase and build stuff incrementally by commits. We are taking a workspace centric approach where multiple agent sessions can happen in a workspace and build based on previous work.
Debatable. Making money is more about structure than being right as per se e.g. short vol is usually right...
The concept overall is basically ok though I think e.g. agents 100% going to be a big thing in finance but it's about man machine synthesis.
it should be easy to self host in docker though.
Excellent! This is exactly what we need!
Coming up next month, adapt this Walmart kayak for your next trip to the moon!
This is slop - the mcp could expose a query endpoint
Obligatory: Brilliant Work. Brilliant.
"We wanted both and couldn't find it, so we built it and open-sourced the whole thing."
\m/ \m/ /m\ /m\