Arcjet, NOAN, and Solid: Three Ways to Stay in Charge of Your Agent
In May, an operator handed an agent a cloud key, a deadline, and one job. Scan a hobbyist network and build an index of it.
0:00 / 7:28
Arcjet, NOAN, and Solid: Three Ways to Stay in Charge of Your Agent
Three of this week's launches are not competing with each other. Each one draws a different boundary around an agent that is already doing real work: what it may do, what it treats as true, and what it may spend. All three reached the board inside the same seven days, and all three exist for the same reason — the agent, not the person, is now the one taking the action.
The short version. Choose Arcjet when your agent can move money, touch customer records, or call an internal system. Choose NOAN when the same question about your own company gets three different answers depending on whether a person, a page, or an agent asked it. Choose Solid when there is a whole job your team keeps failing to get to and you can check the output.
| Tool | Best first job | Setup friction | Time to first value | Small-team price | Integration pain | Practical pick |
|---|---|---|---|---|---|---|
| Arcjet | Enforce what an agent may do at the moment it acts | Add the package and call it in the handler that takes the action, or hand the install to your coding agent; no gateway, proxy, or infrastructure to run 1 | A decision object comes back before the tool runs; local decisions are under a millisecond, about 20–30 ms when the cloud API is needed 1 | $25/month for one person, $299/month for two with a day of log retention, plus $5 per million web requests and $50 per million agent requests; 15-day trial 2 | Their own docs list what the coding-agent hooks cannot do, including the ones that fail open 3 | Agents that move money, touch customer data, or call internal APIs |
| NOAN | Give every page, app, and agent one version of the truth | Make a key and read facts over plain REST; a single command wires the MCP server into Claude Code, Cursor, or Codex 4 | A first request takes minutes; making it your source of truth is a migration, not an install | Free with 1 project, 40 facts, and 150 contacts; the page positions the product at roughly $99–299/month against the CRM, CMS, wiki, and task board it replaces, with extra seats at $79 5 | It wants to hold your customers, tasks, and content, so the hard part is the migration and who owns the facts 6 | Teams whose facts are scattered across a CRM, a wiki, and somebody's head |
| Solid | Hand over a whole job, with the agent holding its own machines and accounts | Describe the job in plain language; the FAQ says you do not set the tools up yourself, you set access and approval rules 7 | Depends on the job; the $20 trial is meant to show what your work costs before you choose a plan 8 | $40, $160, or $640 per month, and the whole amount becomes a balance for AI usage, machines, and anything the agent buys, with no platform fee on top 8 | The balance is also the budget: an agent that signs up for and pays for its own services is choosing how your money is spent 7 | A whole job nobody has time to do, whose output you can check |
Arcjet: the boundary is the action, not the prompt
Arcjet is runtime security that ships inside the code you already deploy. It launched on September 21 and finished the week inside the top five 9. Its pitch is about where a check can sit: a prompt scanner sees tokens, a gateway sees a packet, and a dashboard sees yesterday, but none of them sees the argument about to be passed to the function that moves the money 10. Their own example is that a refund of $12,000 and a refund of $12 look the same from the network.
In practice you wrap the tool your agent is about to call and branch on the answer before it runs: allow, block, redact, or hold for review. Around that, Arcjet detects prompt injection in user input and in what an agent fetched, strips names, addresses, and card numbers before they reach model context, and caps tokens per user, per organisation, and per agent run, so one loop cannot spend a month's budget by touching four endpoints 10.
The install is the part worth noting for a small team. The quick start is an install, a key, and a call in the handler that takes the action, with no gateway, proxy, or Redis to keep alive 1. Their docs claim under a millisecond for local decisions and 20–30 ms when the cloud API answers, and the SDK fails open by default when Arcjet is unreachable, which you can change 1. For Claude Code, Copilot, Codex, and Cursor there is no package at all: you publish a policy in their console and the agent's existing hooks ask before each tool call 3.
Read the checkout page before you commit. One person is $25/month with an hour of log retention; $299/month covers two team members and a day of retention, and usage is $5 per million web requests and $50 per million agent requests, after a 15-day trial 2. A reviewer on Product Hunt who already uses their WAF said it works well and asked for more than the one-hour window on recent requests 11.
The diligence that matters is on Arcjet's own documentation page, which lists where enforcement stops. Claude Code and Copilot hooks fail open on a timeout, a network error, or a non-2xx response, so a slow Arcjet allows the action. Copilot cannot honour a prompt denial at all. Codex's hosted tools skip the pre-tool-use hook, so a tool-call policy never sees them. A recorded event can never block, because the tool has already run — which means a poisoned web page the agent fetched is past the checkpoint. And a hook binds one client, not one person: a personal Claude or ChatGPT login on a corporate laptop is invisible to hooks and to compliance 3.
Choose Arcjet if you are shipping an agent that can move money, touch customer records, or call an internal system, and you want the decision to happen at the action rather than in the prompt. Skip it while your agent only drafts text for a person to approve.
NOAN: approve the fact, then let everything read it
NOAN launched on September 24 under the line "the fact layer for your AI agents" 9. The distinction it is selling is sharp: a fact is something true about your company — approved, versioned, and served from one API — while memory is what an AI picks up along the way that nobody signed off 6. Through the API and an MCP server, the same facts feed any model, app, or agent. Their own site is the demonstration: the pricing page renders live from the fact layer, which is why the plan prices on it are current rather than copy 5.
Setup is the lightest of the three here. Create a key, send one request, and your facts come back as plain JSON, with no SDK and no framework constraint; the API is plain REST with a bearer token, at 200 requests a minute on self-serve plans. Writes are attributed to the key's owner and versioned with history, and agents that propose changes to a fact wait for human verification. There is also a single command,
npx -y @getnoan/wizard@latest, that wires the MCP server into Claude Code, Cursor, or Codex, installs their skill, and hands the seeding of an empty workspace to your assistant 4.The free plan is one project, 40 facts, 150 contacts, and one API key. Paid plans add seats — $79/month each, or $59 on the team plan — and extra projects at $79/month; enterprise adds seats at $125, a $7,500 custom onboarding, and a $1,000/month premium support option. The pricing page puts the product at roughly $99–299/month against the stack it says it replaces: a headless CRM at $25–75 per user, a CMS at $99–300, a knowledge base, a task manager, and the API layer you would otherwise build yourself 5.
Two diligence points. The home page says everything is free except the facts, while the pricing page sells seats and projects — read which plan you would actually land on, because 40 facts is a sample rather than a fact layer. And this is a bigger decision than a seat: NOAN wants to hold your truth about customers, tasks, and content, so the real work is the migration and deciding who approves a fact.
Choose NOAN if the same question about your own company gets a different answer depending on whether a person, a page, or an agent asked it. If your facts already have one home you trust, you have solved this.
Solid: the agent gets a wallet, and the wallet is your plan
Solid launched on September 23 as the top product of its day 12. You describe a job in plain language and close your laptop; the agents work out the steps, connect to the tools, build whatever is missing, check the result, and message you when it is done 7. What makes that possible is the part to think about: they get their own devices — Linux servers, Windows and macOS machines, iPhones and Android phones — their own Google and Apple accounts, and a budget they can spend. Their FAQ is explicit that no prebuilt connector is required and that you do not set the tools up yourself; you supply access and choose the approval rules. Agents keep the fixes that worked, so a correction carries into the next job, and they can pull in more agents or call Codex and Claude Code when a job needs a team 7. There is an API for putting always-on agents inside your product, and an enterprise option that runs in your own VPC or on-premises.
The pricing model is one balance, and it is the whole story. $40, $160, or $640 per month, and the full amount becomes a balance that pays for AI usage, machines, and anything the agents buy — with no platform fee taken out, and $20 of trial credit to find out what your work costs 8. That is a clean deal. It also means the agent determines how your money is spent, from inside your own plan.
There is a documented version of that failure. In May, an agent with a cloud key and a deadline rented five of AWS's large memory-optimised instances to scan a hobbyist network, ran for about twenty-four hours before its operator shut it down, and left a bill of $6,531.30 — with the operator asking that community for donations 13. The Hacker News thread on it landed on the structural point: the account had no spending limit to set, and no cap on instance types 14. That incident is not Solid's, and the comparison is not that Solid repeats it. It is that a place where agents hold accounts and money is exactly where a spend boundary has to exist, and you are the one who sets it 8.
Their answer to that is budgets, approval rules, and a meta-agent that can explain itself: ask what a job cost, ask what is still running, ask what could be stopped. Make that the first job. Spend the trial credit asking for a breakdown before you hand over anything with a price on it.
Choose Solid if there is a whole job your team keeps failing to get to and you can check the output. It is not the tool to hand your production credentials to on day one.
The startup-job picks
For an agent that acts — moves money, touches customer records, calls internal systems — start with Arcjet. It is the cheapest of the three to try at $25/month, and the thing to test is whether the boundary holds on one reversible action, in dry-run mode, before you set it live 210.
For facts your whole team keeps re-litigating, start with NOAN. Stay on the free plan with a handful of real facts, and find out within an afternoon whether one source of truth is the problem you actually have 5.
For a job nobody on the team can get to, start with Solid. Put twenty dollars of work behind one job whose result you can check, then ask the agent what that job cost 8.
What the three share is why they belong in one episode. The interesting shift this week is not that agents got smarter; it is that they are now the ones acting, holding accounts, and being believed. That turns one big question — which model is best — into three smaller ones: what may it do, what does it know, and what may it spend. Buy the boundary your team keeps getting wrong.
One limitation matters. This is a public-source buying screen, not a logged-in trial inside each product. The setup paths, prices, and limits here come from the vendors' own documentation and pages, including the limitation sections they publish about themselves, plus what the discovery boards recorded about when each product launched. Verify runtime behaviour on one real but reversible workflow before you put any of them in front of paying customers.
References
- 1Arcjet get started
docs.arcjet.com
- 2Arcjet pricing
arcjet.com
- 3Arcjet: secure coding agents
docs.arcjet.com
- 4The NOAN API
getnoan.com
- 5NOAN pricing
getnoan.com
- 6NOAN
getnoan.com
- 7Solid
solid.tech
- 8Solid pricing
solid.tech
- 9Product Hunt weekly leaderboard, week of September 21 2026
producthunt.com
- 10Arcjet
arcjet.com
- 11Arcjet on Product Hunt
producthunt.com
- 12Solid on Product Hunt
producthunt.com
- 13
- 14Hacker News discussion
news.ycombinator.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.