Open-source · Apache-2.0 · v0.0.7

Growth on autopilot.Every decision in the open.

Drip runs the whole user-acquisition loop — diagnose, decide, allocate, create — as a team of agents. The decision core is deterministic and auditable; the LLM only explains. Self-host it, fork it, see exactly why before you spend a dollar.

Any of 12 LLMs Meta · TikTok · Tencent · Ocean Engine · Kuaishou Your data, self-hosted
drip · decision card
Meta_Prospecting_v3DTC_US · 7-day window
SCALE +20%$200 → $240/day
CPP
$18
ROAS
3.8×
CVR
2.5%
CTR
1.4%
Freq
1.8
Spend
$200
Purch
11
Headroom
83%
rule chain: 7/8 green · unit economics healthy (CPP<target, ROAS>3×) · sample on the edge → scale the conservative step (+20%), not +50%.
confidence HIGH · revert if CPP > $22 Approve ✓
Rules decide, the LLM only explains 8-signal decision engine Drip-Bench · reproducible shadow → copilot → autonomous
Why Drip

Human strategy. Agent execution.
Nothing hidden in between.

The closed UA agents say “trust us.” Drip shows the work — every scale, pause, and budget move opens its full reasoning.

Rules decide, not vibes

Eight signals → red/amber/green → a deterministic rule chain produces the action, confidence, and guardrails. The LLM only writes the human “why”. Replayable, byte-for-byte.

The whole loop, one command

collect → diagnose → strategize → create → allocate → learn. drip run walks the entire cycle, cross-platform, and reallocates budget from losers to winners within your cap.

Safe by default

A ladder, not a switch: shadow (plan only) → copilot (human approves each write) → autonomous (within a hard budget cap). A person signs off before money moves.

How it works

See it. Decide it. Prove it.

Platform AI won the bidding auction. The new battleground is the decision layer — and that's exactly what Drip opens up.

01 — DIAGNOSE

Every campaign, scored on 8 signals

CPP, ROAS, CVR, CTR, frequency, spend, purchases, headroom. Each goes red/amber/green, the rules fire, and you get a decision card with the full vector + rule chain + confidence.

  • Thin sample → it scales conservatively and caps confidence
  • Open the rule chain to see exactly why
TikTok_Broad_v1unit economics broken
PAUSE$240 → $0
CPP
$41
ROAS
1.4×
CVR
1.2%
CTR
0.8%
rule chain: CPP & ROAS both red — unit economics broken → stop the bleed. Confidence capped LOW (sample below minimum).
02 — ALLOCATE

Budget flows to winners, across platforms

Nobody optimises across walled gardens — that's the open, neutral gap Drip fills. Freed budget from pauses flows to the scalers in proportion to value, inside your daily cap.

  • Cross-platform, vendor-neutral
  • Value-weighted (ROAS by default, plug your LTV model)
meta · Prospect_v3$500SCALE
tiktok · Prospect_v3$500SCALE
meta · Broad_v1$0PAUSE
tiktok · Broad_v1$0PAUSE
03 — PROVE

The first reproducible UA-agent benchmark

10 hand-curated decision cases, a three-part rubric, a pluggable agent interface. Score any model — yours, ours, a competitor's — and every run writes a bundle anyone can re-run and diff.

  • No more unverifiable “70% lower CPA”
  • Open leaderboard — win or lose, we publish ours
$ drip bench run --agent claude
drip-bench v0 · 10 cases

▸ 001 sample on the edge  92/100
▸ 002 slow bleed         88/100
▸ 003 budget reallocation 81/100
▸ 004 creative fatigue    95/100


  total 851 / 1000
  bundle → benchmarks/runs/…
04 — EXECUTE

It writes — to Meta and China, gated

drip apply pushes scale / pause to Meta · 腾讯 · 巨量 · 快手 (auto-routed); drip watch guards pacing intraday; drip autopilot runs the whole loop, signal-routed behind a circuit breaker. shadow → copilot → autonomous, every write capped + audited.

  • Cross-border: Meta · TikTok · 腾讯 · 巨量 · 快手
  • Autonomous — but it halts on anomalies or write failures
$ drip autopilot --mode autonomous
situation: scaling → scale-winners → allocate

  meta  Prospect_v3  SCALE  applied ✓
  腾讯  Prospect_v3  SCALE  applied ✓
  巨量  Broad_v1     PAUSE  applied ✓

  ⛔ breaker ok · audit → writes.jsonl
Open vs closed

Why a transparent agent wins

We're not claiming Drip beats them on raw performance today. We're claiming it's the only one you can audit, fork, and run in your own environment.

 
Sett · Kohort · GrowthGPT
Drip
Code
Closed
Apache-2.0
Why a decision happened
“Trust us”
Signal vector + rule chain + replay
LLM
Vendor-locked
Any of 12 + local
Cross-platform
Per walled garden
Neutral, optimises across
Your data
On their servers
On yours, self-hosted
Evaluation
Marketing claims
Drip-Bench, reproducible
Price
$99–$999+/mo
Free

See every decision before you trust it with a dollar.

Open the live console — no install, no signup.