The browser layer for AI agents.
Agents reason brilliantly but break the moment they touch the live web. Monopea is the execution layer they call to act — a token-efficient browser API that turns a goal into deterministic, replayable action.
Indexed DOM state · Deterministic skill replay · Human-in-the-loop
POST /api/v1/run
Authorization: Bearer ab_live_…
{
"url": "https://app.acme.com/invoices",
"prompt": "export last month as CSV",
"success": { "urlIncludes": "/exports" }
}
→ 200 { success: true, path: [3 steps], … }The web wasn't built for agents to act on.
LLMs are great planners and terrible operators. The gap isn't intelligence — it's the interface between a goal and a live, hostile, ever-changing page.
Raw HTML burns context
A single modern page is 50k+ tokens of markup, scripts, and noise. Feed that to an agent every step and you pay for — and reason over — almost nothing useful.
Dynamic DOMs break scripts
Class names rotate, structure shifts, selectors rot. Static scrapers shatter the moment the page they were written against changes.
Every run re-reasons
Multi-step flows make the model think through the same path on every single run — slow, non-deterministic, and expensive at any real volume.
Goal in. Deterministic action out.
Three stages take an agent from a free-form prompt to a repeatable skill it can run for free, forever.
Indexed state
Every page is compressed to a token-efficient, numbered map of what an agent can actually act on — not 50k tokens of HTML.
[1] Button "Log in"
[2] Input "Email"
[3] Link "Forgot password"Skill compilation
Discover a goal’s path once with the planner, minimize it to the shortest reliable route, then replay it deterministically at zero LLM cost.
discover → 7 steps
minimize → 2 steps
replay → 0 LLM callsHuman-in-the-loop
Hit an authorized approval or MFA step and the run pauses, hands off to a human, and resumes exactly where it left off — by consent, never by bypass.
status: paused
reason: approval
POST /sessions/:id/resumeEverything an agent needs to act — and nothing it shouldn't.
A complete execution surface: efficient perception, deterministic action, safe handoff, and the tenancy and audit primitives to run it in production.
Token-efficient DOM
Pages compressed to an indexed action map — orders of magnitude fewer tokens than raw HTML, with the same actionable surface.
Deterministic skill replay
Compiled skills are descriptor-based and survive DOM rotation, replaying the minimal route with no model in the loop.
Human-in-the-loop handoff
Pause on an authorized approval or MFA step, hand to a human, and resume the run — sessions survive the wait.
MCP + LangChain
Drop into Claude, LangChain, or AutoGen with one line. Thin adapters over the same REST engine — no logic reimplemented.
Per-tenant API keys
Scoped, hashed keys per tenant with Row-Level-Security isolation. The raw key is shown once, then only its hash is stored.
Full run audit log
Every run records its goal, path, success condition, and authorization reference — an auditable trail for every action taken.
Automates only authorized targets.
Monopea is built for first-party sites, operator-approved partner automation, internal RPA, and accessibility. It is explicitly not an anti-bot evasion service — no WAF tuning, no CAPTCHA bypass, no removing the allowlist.
Per-tenant allowlist
Every execution call is gated by an allowlist of hosts you’re authorized to automate. Unlisted targets return 403 — no exceptions.
Signed authorization
Each authorized host carries a signed authorization reference — scope, window, and rules of engagement — stamped into the run record.
No evasion, ever
We build both the detector and the execution layer. We automate with consent; we do not defeat anti-bot defenses for hire.
One line to add the browser layer to your agent.
MCP, LangChain, AutoGen, or raw REST — the same consent-gated engine behind every one. No discovery or replay logic to reimplement.
# expose the engine over MCP (stdio)
npm run exec:mcp
# tools: run_goal · compile_skill · run_skill · list_skillsimport { makeMonopeaTools } from "monopea";
const tools = makeMonopeaTools()
.map((t) => new DynamicStructuredTool(t));curl https://monopea.com/api/v1/run \
-H "Authorization: Bearer ab_live_…" \
-d '{ "url": "…", "prompt": "…", "success": {…} }'Action-credits. You pay for what your agent does.
A read costs 1 credit. A solved semantic flow costs 10. Compile it once into a skill and every replay drops back to 1 — discovery is the only premium you pay.
- Pay for actions, not seats
- Free tier with 1,000 credits / month
- Compiled-skill replays cost 1 credit
- Usage-based Scale tier — no surprises
Give your agent a browser it can trust.
Start with 1,000 free credits. Compile your first skill in minutes and replay it for the cost of a single read.