BotRampart

Open-source bot defense, self-hosted

Price the request, not the visitor.

BotRampart never decides who is a bot. It runs inside your own app server, gives every session a budget and every path a cost, and asks heavy users to pay with work: a proof-of-work, then a human check, then a wait. No proxy in front of your site, no device fingerprint, and the record of every decision stays in your own bucket.

In design. Nothing is built or installable yet; this site previews the plan, and every number from someone else is sourced.

Two sessions, one policyeach line = 10 s

A person browsing the catalogue

tier 1 · signed out

credits300
  1. GET /products/42−1
  2. GET /products/42/reviews−1
  3. GET /api/search?q=boots−5
  4. refill, 20 a minute+7
  5. GET /products/77−1
  6. no challenge after tier 1ok

A script working through search

60 searches every 10 s

credits0
  1. proof-of-work solved+300
  2. 60 × GET /api/search−300
  3. over budget: human check403
  4. check passed (paid solver)+3,000
  5. repeat pass, grant halves+750
  6. grants spent: throttled429
Simulated with the proposed default policy: search costs 5, refill 20 a minute, grants 300 then 3,000, halving on repeat.

01 · Why now

Telling bots from people has become a guess you can buy your way past.

More than half of web traffic is automated, the puzzles meant to separate people from machines are solved by machines, and the IP addresses that used to identify an abuser now rent by the gigabyte. A verdict of "bot" or "human" is cheap to fake. A budget still has to be spent.

53%

of the web traffic Thales observed in 2025 was automated, by its annual count.1

100%

of reCAPTCHA v2 image challenges were solved by a published model in a 2024 study.2

$1–$2.99

per 1,000 solved reCAPTCHA v2 challenges at a solver service; people and models do the clicking for you.3

65.3%

of 21,491 domains in DataDome's 2026 test stopped none of its 10 test bots. A vendor's test, and it favours the vendor's case.4

"Is this a bot?" A classifier decides, and you defend the answer.

Every false positive is a customer you blocked and cannot explain. Every pass is a clearance that can be resold. Solving the challenge once buys the session unlimited throughput.

"What has this session spent?" Arithmetic decides, and anyone can check it.

Ordinary browsing stays under the refill rate and never sees a challenge after the first invisible one. Heavy use runs out of credit, and each challenge buys a bounded, shrinking amount more. Nobody is labelled; everybody is metered.

02 · How it works

Three loops at three speeds, and only the fastest one is in the request path.

The backend SDK decides each request in memory, with no network call. A five-minute loop looks for patterns no single session shows. People stay in charge of the policy.

Microseconds · in your app

Per request

The backend SDK checks the signed session cookie, charges the path's cost to the session's budget and runs your policy rules. Allow, challenge, throttle or block, before your handler runs.

Knows
This session and the last synced policy
Calls
Nothing
Every 5 minutes · in Core

Per window

Core turns the last window into named bands ("cost x6 over baseline", "new sessions over 60%") and asks a decision model typed questions. The answer becomes fields your rules can use. It can raise a price; it can never block.

Knows
Aggregates, never strings a visitor wrote
First build
Shadow mode: answers are logged, not enforced
Human time · in Slack

Per question

An agent reads the decision log in your bucket with read-only access and answers "why was this person challenged?" with the rule, the numbers and the reference. It proposes changes; a person approves them.

Knows
Everything in your bucket, read-only
First build
Explanations and a daily digest
Where BotRampart sits Visitors reach your app through your existing CDN. The backend SDK inside the app verifies and charges each request with no network call. It syncs counters with Core every few seconds. Core asks a decision model about bucketed state every five minutes and writes everything to your own object storage, which the agent reads to answer questions in Slack. Visitor browser SDK, PoW Your CDN / WAF unchanged, floods Your app server backend SDK: verify cookie, charge budget, run rules no network call Your handler req.br has the decision BotRampart Core sync, challenges, policy, 5-minute state builder counters, every 5–10 s policy, directives Decision model named bands only, every 5 min, cached Your bucket R2, S3 or MinIO Parquet + JSON Agent DuckDB, read-only answers in Slack
Where BotRampart sits. Orange is the request path, which gains no extra hop. Green is what stays in your infrastructure. The decision model is the only outside call, it sees named bands rather than visitor data, and it can be a model you host yourself. Our design.

03 · Control

It runs where your app runs. Your traffic data never leaves.

Most bot defense is a service your traffic passes through: a proxy, an edge network, or a module that asks the vendor about every request. BotRampart is software you run.

  • No proxy in the request path. The SDK is a framework filter in your app. Your CDN or WAF keeps handling floods, as it does today.
  • Verified offline. The session cookie is signed with your key and checked in memory. If Core is unreachable, tier 1, budgets and rules keep working.
  • Your bucket, open formats. Rollups, decisions and every model answer are Parquet and JSON in your R2, S3 or MinIO, readable with DuckDB without asking us.
  • No fingerprint ID. One server-set session cookie, an optional hashed login ID, and IP only as a network bucket. No canvas, font or audio fingerprinting in any mode.
  • Open source. SDKs under Apache-2.0; Core and the agent under AGPL-3.0 with a commercial licence for those who need one.

04 · Explains itself

Ask why. Get the rule, the numbers and the reference.

Support engineer 10:31

@BotRampart why was user 4217 challenged around 10:12 today? They say they were just searching.

BotRampart 10:31

Session BR-7F3K2Q got a human check at 10:12:41 from rule over-budget, policy v12. In the five minutes before, it spent 1,240 units, 96% on search at 5 units a call. A signed-in session regains 300 units in five minutes. Risk score 1; no model directive applied. The check was passed in 4 seconds and the session carried on.

query explain_decision · 3 rows · window 10:07–10:12 · policy v12

Show timelineReplay with draft policy

Illustrative: the answer format from the design, not a real user

Every non-allow decision is logged with a reference shown to the visitor, so support can look it up. Because each decision is a rule, a budget or a recorded model answer, an explanation is a lookup rather than a guess.

05 · Who it is for

Built for teams who cannot, or will not, route their traffic through a vendor.

And honest about who should use something else. If a CDN in front of your site is fine with you, the incumbents already do a lot for free.

A good fit

  • Engineering-led SaaS, shops, marketplaces, travel and ticketing sites with roughly 5 to 200 million requests a month, scraping, fake sign-ups or credential stuffing, and no security team.
  • Teams priced out of enterprise bot management, whose entry plans start in the thousands of dollars a month.5
  • Regulated and data-sovereign teams: fintech, health, government suppliers, EU residency rules.
  • Apps on private networks or behind several CDNs, which cannot sit behind one vendor's edge.

Probably not for you, yet

  • You already run behind Cloudflare and are happy: its free plan includes Bot Fight Mode and Turnstile is free with unlimited challenges.6
  • You need volumetric DDoS protection. BotRampart refuses requests in your app; floods belong at the edge.
  • You need mobile apps or partner APIs protected now. The first build covers browsers only.
  • You need something in production this quarter. BotRampart is in design; the comparison names what to use today.

06 · Status

Where it stands, plainly.

The design is written, the interface is specified and the first build is scoped. Durations are planning assumptions for a team of two or three engineers, not estimates from measured work.

Now

Design

Architecture, OpenAPI spec, policy format and threat model. Nothing to install.

Phase 0 · 2 weeks

Decide

Hosting spike, decision-model bake-off, proof-of-work format measured on phones, licence and trademark.

Phase 1 · 8 weeks

First build

Node SDK, tiers 0 to 2, budgets, Core on Cloudflare or one container, decision log, agent explanations in Slack.

Later

Complete the loops

Model answers enforced, log pack, Python and Java, console, step-up methods, an open benchmark, cloud.

See what the first hour would look like on your traffic.

The preview walks through setup in observe mode and ends on what BotRampart would have shown you. It runs in your browser and stores nothing.

Sources

Every figure from someone else on this page, with the words it rests on. Checked on the date shown; the full ledger is claims.yaml.

  1. 2026 Thales Bad Bot Report (PDF). Automated traffic accounted for 53 percent of all observed web traffic in 2025. Of that, 40 percent was generated by bad bots, up from 37 percent the year before. Checked 2026-11-18.
  2. Breaking reCAPTCHAv2 (arXiv 2409.08831). Our main result is that we can solve 100% of the captchas, while previous work only solved 68-71%. Checked 2026-11-18.
  3. 2Captcha: Captcha prices. $1 — $2.99 Checked 2026-11-18.
  4. Full bot protection drops to 2.4% of popular websites, DataDome finds (PPC Land). Of those domains, 65.3 percent let every one of the company's 10 test bots through without a block or a challenge. Checked 2026-11-18.
  5. PricingSaaS: DataDome pricing 2026. DataDome pricing starts at $3,830/mo on Essentials. Checked 2026-11-18.
  6. Cloudflare docs: Turnstile plans. Unlimited challenges (traffic or verification requests) Yes Yes Checked 2026-11-18.