TradingAgents is a research agent with more than a hundred thousand stars, and its release notes mention proxy support. If a data pull is being refused, that sentence is probably not about your problem. This page separates the two meanings and then goes source by source. We read release 0.5.0 on 20 September 2026.
A note on scope: this is about data collection and network behaviour. There is no trading advice here and no claim about what the agent produces.
Two meanings of one word
In this codebase the word proxy means an interface in front of the model. The code says so directly. A custom base URL on the standard provider is "a proxy, gateway, or local server". Such a server speaks only the older chat interface, so the newer one is switched off when you use it. That is worth knowing, and it has nothing to do with networking.
If you want the model calls themselves to leave through your line, the lever is elsewhere. The configuration forwards http_client and http_async_client straight to the model library. Hand it a client you built with a proxy.
Which call needs what
A search of the whole release for the word proxy finds five files: the readme, the changelog, the command line utilities, the model client and its test. Not one data source. So the data side is answered per source, and the sources are not alike.
| The call | How it is fetched | Does an address help? |
|---|---|---|
| The model | your own client, or a gateway URL | yes, via http_client |
| The forum feed | the standard library, with an identifier | yes |
| The filings service | requests, with a required contact address | no, identify yourself instead |
| A keyed market vendor | an API key with a daily quota | no |
| The optional market library | its own session or the environment | yes, and it has its own page |
Where a proxy is the wrong tool
The filings service is the clearest case in this whole series. It needs no key and no account. What it requires is a User-Agent that identifies you, with a contact address. The project's own comment records the experiment. A client name alone, or with a project URL, gets a 403. One with an address gets a 200.
So set your own identifier, with your own name and address, so they can reach you about your traffic rather than the project. Hiding behind a different address is the wrong direction here. We would rather say so than sell you something for it.
The keyed vendor is the second case. Its limit is your plan, counted per key and per day. The code even recognises that vendor's own wording when it refuses. No address changes that number.
The source where an address does help
The forum feed is a public page fetched with the standard library. So the usual environment variables route it, and the address is visible to the other side. In our own paired test that site answered differently from a residential address than from our server. A hard refusal became a script check.
One thing to leave alone while you are there. The fetch already handles being rate limited properly. It honours the server's own retry delay exactly, and jitters only its own fallback, so parallel runs do not retry in lockstep. That is better behaviour than most agents have. Do not replace it with a proxy and a faster loop.
Which proxy type fits it?
Residential, and only for the fetched sources. That is a small part of this program's traffic, so the honest recommendation is a modest one. A per gigabyte line costs very little here, because feeds, filings and numbers are text.
Buy it for reliability rather than for volume. One address held for a run suits this workload, so a sticky port fits. Write a retry anyway, because a sticky address can change early when its device leaves the network. The residential proxies page lists the plans, and the plan API manages allowed addresses from code if you would rather hold no password.
What breaks
- The newer model interface is off. You pointed the provider at a custom base URL. That is expected.
- The filings service returns 403. Your identifier has no contact address.
- The market vendor says you are out of calls. That is the key and the plan.
- The forum feed goes quiet. Check for a refusal being handled by the built in backoff before you change anything.
- Model calls still leave from the server. A gateway URL is not a network proxy. Pass your own client.
What this page does not cover
We read the release as text and did not run the agent. So we did not measure how often a source refuses, and we did not test the environment variables through its data layer. We read four data modules closely and the rest by name only. This project releases roughly monthly and shipped this version two days before we read it. The vendors in play are configurable too, so both the version and your settings matter. We will read it again by 20 October 2026.
Where to go from here
Proxies for yfinance covers the optional market data library in detail. Proxies for DeerFlow covers another research agent where one tool takes a proxy and the other does not. Proxies for Agent-Reach covers the same split for social sources.
Sources
- The release note mentioning proxy support (README), what that means in the model client and the passthrough list for your own client (tradingagents/llm_clients/openai_client.py), the forum fetch and its backoff (tradingagents/dataflows/reddit.py), the filings identifier rule (tradingagents/dataflows/sec_edgar.py), and the keyed vendor's quota handling (tradingagents/dataflows/alpha_vantage_common.py). TauricResearch/TradingAgents, release v0.5.0 of 18 September 2026, read 20 September 2026.
- Our own research on the optional market data library, 19 September 2026, kept in the research folder of that page.
- Our paired address test of 19 September 2026: 16 URLs, plain requests, two runs from our server and two through a residential line of our house plan. Raw output is kept in the research folder of our OpenClaw page.
- Plans, allowed addresses and sticky sessions. HProxy documentation, hproxy.com/docs, 20 September 2026.
