The systems we ship keep working while nobody is watching. Guardrail sits in the request path on the machine that runs them, so a run that turns expensive is stopped at the proxy instead of explained on next month's invoice.
01 / The partner
NEAT builds Guardrail, a local-first spend control layer for AI coding agents. It runs as a proxy on localhost:4000, between the tool and the provider, and every call passes through it byte for byte. It relays the login the developer already has, so no API key changes hands and there is nothing new to vault or rotate.
What it reads is deliberately narrow: model, token counts, cache behaviour, cost and timing. Prompts, responses and code never reach its logs or its database. That is the trade that makes it usable on a real engineering machine, full visibility of the spend and none of the content.
A cap on dollars, requests, tokens or requests per minute. Over budget the call is never sent, so it costs nothing and the tool gets a clean provider-shaped error instead of a crash.
The cache write tax on a first turn, the reasoning tokens nobody sees, the tool definitions loaded on every call. Attributed per request instead of buried in a monthly total.
A SQLite file in the developer's own home directory, bound to localhost. Nothing is uploaded unless cloud sync is switched on deliberately.
02 / The problem
An operator that acts on its own does the useful work while nobody is looking at the screen. That is the whole point of it, and it is also where the money goes. A retry loop overnight is a real charge on a real card by morning, and a sub-agent loop on a flat plan can eat a five-hour window in twenty minutes and lock a developer out in the middle of the day.
The tools that exist for this mostly report. Three gaps show up every time.
Provider billing is a rear-view mirror. By the time a number looks wrong, the money left the account weeks ago.
A token counter watches the meter run. It has no opinion about when to stop, and no way to act on one.
A threshold alert fires after the spend it is warning about. A budget nobody can enforce is a preference.
03 / The integration
Guardrail goes in front of the agent instead of alongside it, which is the only place a budget can be enforced rather than observed. Four pieces do the work.
The tool is pointed at the local proxy and keeps its own login. Nothing about the agent, the repository or the workflow changes, which is why it can go on a client machine the day it is needed.
English
guardrail start · proxy on localhost:4000Every request is checked against the budget for that project and session first. An over-budget call is never sent upstream, so it is never billed and never counted against a plan window.
policy: DENY · upstream calls: 0 · cost: $0.00Spend is tied to the session and the project rather than to an account total, so the expensive path is the thing that shows up, not the whole month averaged into one number.
per session · per project · per requestSpend pace and a projected month-end total from the runs already made. That is the number a client asks for before approving the next phase, and it stops being a guess.
pace: live · projected month end04 / The path
A single unattended agent run, from the command that starts it to the number a client sees.
One command puts Guardrail in the path. The existing Claude Code or Codex login is relayed untouched, so no credential moves anywhere.
Model, tokens, cache behaviour and cost are read from the metadata while the request body is forwarded byte for byte.
The policy for that project runs pre-send. Under budget the call goes. Over it, the call stops on the machine it started on.
Blocking ends the loop without killing what is in flight, so the failure is one stopped request rather than a broken session and a lost afternoon.
Live spend by session and by project, with a projected month-end from the runs already made, available while the work is still happening.
05 / The partnership
We ship AI systems that act for real businesses and then keep running them. The question a client asks straight after can it do the work is what stops it costing more than it saves, and until now the honest answer was a dashboard and a promise to watch it. Guardrail turns that into a limit the system enforces on itself.
The work runs both ways. Guardrail goes in front of agent runs that operate all day on production paths, which is where the awkward cases live, and the systems we deliver get a spend boundary a finance team can sign rather than a report they read afterwards. We are a certified NEAT partner, so clients get the product, the deployment and the people who run it in one place.
Their site has the current product detail, the supported tools and the free local install.
Book a free 15-minute call. We will help you identify the highest-leverage automation, API integration, AI agent, or internal system to build first so your team can move faster with less manual work.
© 2025 Bles Software. All Rights Reserved.