Talorys is a free, MIT-licensed personal AI assistant by a developer who goes by rociiu on GitHub. One command, npx create-talorys@latest, deploys chat, memory, tasks, notes and reminders into your own Cloudflare account on the Workers free plan, and the chat model is glm-4.7-flash running on Cloudflare’s Workers AI. The repo was created on October 10 and the code I read is version 0.1.5, so it is three days old.
In this post I go through what “free AI agent” means in numbers: what a Cloudflare Neuron is, how many chat turns 10,000 buy per day, and what happens when they run out. Then I check what the repo’s own spending guardrails guarantee, and whether “self-hosted” is the right word. My verdict: the free tier is real but small, the default cap of 200 requests a day protects you less than it looks like, and I would call this “self-deployed”. I did not deploy it, because that needs a Cloudflare login. What I did run is the repo’s own budget tests and a calculator.
Here is why I care. At Trellis I want an assistant that answers an Amazon seller’s questions about their listings, and the first question is what it costs when the seller asks 400 things on a Monday. I wrote before about a budget cap that did not cap, and a free allowance is a design constraint, not a business plan.
What is a Cloudflare Neuron, and how many chat turns is 10,000?
A Neuron is Cloudflare’s billing unit for AI. The Workers AI pricing page says it represents “the GPU compute needed to perform your request”. The free allowance is 10,000 Neurons per day, reset at 00:00 UTC, which is 8 pm in Toronto right now. On the Workers Paid plan, anything above that costs $0.011 per 1,000 Neurons, so the whole free allowance is worth $0.11.
For glm-4.7-flash the page lists 5,500 Neurons per million input tokens and 36,400 per million output tokens. Output is 6.6 times pricier than input, token for token. That is the number that matters, because a chat agent sends a lot of text in and gets little back, but the little it gets back is the expensive part.
Now the arithmetic, which I did in Python and you can do on a napkin. These token counts are my assumptions for a turn, not measurements from a running Talorys.
- A light turn, 2,000 tokens in and 500 out: 2,000 × 5,500 / 1,000,000 = 11 Neurons in, 500 × 36,400 / 1,000,000 = 18.2 out. Total 29.2, so 10,000 / 29.2 gives about 342 turns a day.
- A full-size turn at Talorys’ own defaults, 6,000 in and 1,024 out: 33 + 37.3 = 70.3 Neurons, so about 142 turns a day.
So 140 to 340 turns, depending on how much memory and history the agent drags into each call. One person in the Show HN thread wrote that they spent 500 Neurons in a few minutes of back and forth, and I cannot check that, but it fits my numbers.
What happens when the free Neurons run out?
The chat stops and everything else keeps working. The README says tasks, notes, memories and reminders keep going, and chat resumes after the daily reset. I checked that in the code: apps/agent/src/ai/provider.ts catches Cloudflare’s error 4006 (“daily free allocation exhausted”) and turns it into an AI_QUOTA_EXHAUSTED message that says the same thing.
That is the right design, franchement: the app does not need the model to show you a task list. Simple reminders and task digests never call AI at all, says the README.
The catch is the plan type. On the free plan, hitting the allowance is a wall. On Workers Paid, it is not a wall, it is a meter running at $0.011 per 1,000 Neurons. A commenter in that same thread says they were billed for Neuron usage on a paid account that they thought was covered, and that support ignored them. One unverified report, so read it as a reason to check your dashboard.
Does the 200-requests-a-day cap protect a free AI agent?
Not by itself, which I did not expect. The defaults, from constants.ts: 1,024 max output tokens, 6,000 max context tokens, 6 tool calls and 4 agent steps per request, 6 scheduled AI runs a day, and maxAiRequestsPerDay: 200. The README says every model call counts toward that last one, reasoning steps and history summaries included. Hard ceilings: 4,096 output tokens, 32,000 context, 2,000 requests a day.
Now put the two together. The cap counts calls. Neurons are charged by tokens. Two hundred light calls cost 200 × 29.2 = 5,840 Neurons, which is fine. Two hundred full-size calls cost 200 × 70.3 = 14,055 Neurons, which is 40 percent over the free allowance. The cap and the allowance cross at about 142 full-size requests.
Mettons you are on a paid Workers account and you raise the cap to the 2,000 ceiling. Full-size requests all day would be about 140,500 Neurons, roughly $1.44 above the free part, which is not scary. The cap is a counter of calls, not a budget, and the Usage panel only shows local estimates; the README sends you to the Cloudflare dashboard for the exact figure.
To be fair to the author, he knows. Issue #30, “Track token provenance and estimate model-specific AI consumption”, was opened October 13 and is about exactly this. If you are on a paid account in the meantime, I would drop the cap to 100. At full size that is 100 × 70.3 = 7,030 Neurons, under the free line even in the worst case.
I also ran the repo’s own ai-budget test file with Vitest after npm install: 3 passed in about 12 seconds. It uses a fake model that counts calls, so it shows the counter stops a turn at the limit and counts a history summary as a call. It says nothing about Neurons. Wrangler also warned that AI bindings reach remote resources and may cost money even in local dev.
Is Talorys really self-hosted if it lives in your Cloudflare account?
No. Me, I would call it self-deployed. The data sits in one SQLite-backed Durable Object in your account, the author says there is no telemetry and nothing goes to him, and the owner password is hashed locally before it becomes a Cloudflare secret. That is real ownership of the data.
But you cannot unplug it. The chat runs on Cloudflare’s GPUs, the README itself says Cloudflare processes your messages and the memories included in each prompt, and the whole stack (Pages, Workers, Durable Objects, Workers AI) is Cloudflare-only. The README says quotas “can change”, and your agent changes with them. The label argument in that thread ran long, and I do not think it matters as much as the practical question: can you leave? Here you can, in part. Settings has a “Download backup” button that exports conversations, memories, tasks, notes and projects as one JSON file. The code does not run anywhere else, though, so you take the data and lose the app.
Should you deploy it?
I would not call this a finished product, and neither does its version number. It is 0.1.5, three days old, with 7 open issues and pull requests on the day I checked, one of them (#1) a request for MCP server support. I have not used the chat, so I have no opinion on whether the agent is any good at remembering or planning.
Deploy it if you already have a Cloudflare account, want a personal assistant to poke at, and will read the TypeScript. The Worker has no public URL and sits behind a service binding, which I like. Skip it if you are on Workers Paid and will not look at the dashboard, or if you want web search from your agent. A commenter asked for that and it is not there yet.
What I do not have is the number I care about: what 50 real turns cost in the Cloudflare dashboard against what Talorys’ own Usage panel says. I need a throwaway Cloudflare account for that, and when I have one I will run it.
