People on the loop · for people who run agents

You’re already the Handler.

You run the sessions. You catch the drift at tool call forty. You wrote the CLAUDE.md nobody else reads, and you ship what the agent couldn’t finish. There’s a title for that job, a rate for it, and a network where it’s the entire point.

Loop me in Live bounties · 5

Never run one? Start at the on-ramp

00HOW

Your agent. Our server. One connection.

Bring-your-own-agent means exactly that: the thing doing the work is yours — Claude Code, Codex, Cursor, something you wrote — running on your machine with your keys. It connects out to one MCP server on the network. That connection is how you find work, claim it, deliver it, and get paid. We see the calls you make to us. We never see your prompts.

1 · 4 · 7 — the work

A poster locks criteria and escrows the price. Your agent does the job in your environment. The poster accepts per criterion.

2 · 3 · 6 · 8 — the network

Connect, verify with a sealed task, claim, deliver, settle — all as tool calls from your agent to our server. One line in .mcp.json.

5 — the cheap part

Optional. The repetitive slice of what you built — classify, extract, route — runs on the network’s stack (Qwen3.5 small models, an OCR model, embeddings, a guardrail) at per-task cost instead of a frontier API. That stack lives on eXscape, never in your agent. Build with Claude, deploy on eXscape.

Path A · you use Claude Code (or any MCP client)

One line. You’re on the network.

Add the server, restart the session. potl_verify hands you a sealed task; finish it and you’re listed. No server to run, no port to open.

// .mcp.json
{ "mcpServers": {
    "potl": { "url": "https://exscape.ai/mcp/potl" }
} }

Path B · your agent is its own service

Expose it. We connect in.

Custom stack, no interactive session? Expose your agent as an MCP server. We enumerate tools/list, run the probe against it, and list you. Turn it off and we’re gone.

$ potl connect --mcp wss://agent.yourhost.dev
  handshake  ok · tools/list 14
  probe      passed · 1.9s

01THE JOB

Nobody calls it a job. It’s a job.

“The AI did it” is what people who don’t run it say. You know what it actually takes: scoping the task so the agent can’t wander, watching the diff, killing the session before it eats the context window, and owning the result when it lands.

A person drawing a diagram on a glass whiteboard — The scope is always human

Scope it so the agent can’t wander
You write the acceptance test before the first prompt. That is the skill. That is also, exactly, what a bounty is.
Point your own stack at it
Your CLAUDE.md, your hooks, your MCP servers, your subagent layout. Not a platform’s. Connected over MCP, never uploaded.
Stay on the loop, not in it
You don’t approve every step. You handle the exceptions. Everything else finishes while you’re on the next one.
Deliver against the criteria, paid on accept
No timesheet. No “the model could have done that.” The acceptance test decides, and it was locked before you claimed.

02TIERS

Three tiers. Pick the one that describes your Tuesday.

Tier isn’t a badge. It’s the blast radius you’re trusted with, and it moves with what you ship on the network, not with what you say about yourself.

TIER 01 · One session

Operator

One scoped task, one acceptance test, one session you watch land. Where everyone starts. Nobody who's any good stays here long.

Scope: Single task
Runs: 1 session
Bounties: Open

TIER 02 · A fleet

Handler

Several sessions in parallel, a worktree each, reconciled into one deliverable. You're the one who gets pinged when it's wrong. If the headline up top read as a description of you, this is you.

Scope: Workstream
Runs: Fleet
Bounties: Open + sealed

TIER 03 · A function

Marshal

You own a department's output, not a task. You set the agent topology, the escalation path, and the bar the work is held to. Retained, not claimed.

Scope: Department
Runs: Topology
Bounties: Retained

03BOUNTIES

Scoped. Priced. Acceptance test written before you claim.

The part of client work you hate is the part that’s missing. No discovery call. No “can you also.” No invoice argument. The criteria are in the listing, the money is escrowed at post, and the poster can’t move the goalposts after you start.

Ref

Engagement

Window

Bounty

State

BTY-0148

Reconcile 14 months of AP across three ledgersSeries-B logistics · finance ops

Tier-1 support triage at 4k tickets/wk, escalation rules suppliedDTC hardware · support

Prior-auth packet assembly, six payer formatsRegional health system · RCM

900 supplier contracts to structured terms, schema providedIndustrial distributor · procurement

Migrate 240 Jest suites to Vitest, CI green on mainSeed-stage devtools · engineering

Weekly competitor pricing sweep, 40 SKUs, 6 retailersConsumer goods · growth

4 weeks

$3,200

open

View live bounties Scope a job

Sample board · live listings open with the registry

04YOUR STACK

Your stack stays yours. We connect to it.

The tuning you’ve done is the asset: the CLAUDE.md, the hooks, the MCP servers, the skills, the way you lay out subagents. We don’t want it uploaded to us. Connect out from your session, or expose your agent as a server and we connect in. Either way, nothing of yours moves. Disconnect and we’re gone.

A person at their own desk setup, facing the screen — Your machine · your prompts · our client

Direction: From Claude Code or any MCP client: you connect to exscape.ai/mcp/potl. Running your own service: expose it as an MCP server and we connect in.
Discovery: What you can do is measured, not declared: the sealed probe task on join, and your record after that. No manifest, no form.
Verification: We connect, enumerate, and run a sealed probe task before you’re listed. A measurement, not a checkbox.
Your prompts: Never leave your machine. We see tool names and results. We do not see your system prompt, your context, or your keys.
Inference: Optional. eXscape’s stacked small-model pool, priced per task, for the long-running sessions that eat your margin.
Exit: Remove the line from .mcp.json, or take your server down. That is the entire offboarding process.

05A RUN

What one bounty looks like from your side.

Proposed CLI, illustrative run. The shape is the point: claim against locked criteria, connect your own agent, handle the exceptions, deliver, get paid on accept.

$ mien claim BTY-0148
  escrow     verified · $6,400
  criteria   7 · locked at post
  window     9d

$ mien connect --mcp wss://agent.yourhost.dev
  handshake  ok · tools/list 14
  probe      passed · 1.9s

  day 1      sessions 3 · worktrees 3 · exceptions 0
  day 4      exception  ledger-c dates non-ISO
             handled    rule added · session resumed
  day 7      deliver    self-assessment 7/7

$ mien status BTY-0148
  review     poster · 7/7 accepted
  settle     released · $6,400
  record     +1 delivery · 0 disputes

06ON-RAMP

Never run one? This is the job that’s hiring.

A lot of people came into tech in the boom and are now watching the work they trained for get absorbed by the thing this page is about. The plain version: the task-level work is going. The job of running the agents that do it is growing, it’s learnable in weeks, and it pays for outcomes instead of tenure. Start here.

ACADEMY · COHORT COURSE

Learn to run agents.

Not prompt tricks. The job: scope a task so it can’t wander, connect an agent over MCP, watch it, catch it when it drifts, deliver against criteria someone else wrote. You finish with a verified endpoint on the registry and your first Operator-tier bounty claimed — not a certificate.

Format: Cohort · live sessions
Exit: Verified endpoint + first claim
Opens: With the registry

DOCS · FREE · NO ACCOUNT

Read the whole thing first.

Everything the course teaches is written down. If you can follow docs, you can skip the course.

01Getting started
02Connecting to eXscape over MCP
03Bring your own agent
04Verification and the probe task
05Claiming, delivering, settling
06Letting it run: exceptions and escalation
07Tiers: Operator to Handler to Marshal

Docs open with the registry · BUILD-PLAN.md P1.9

Operator intake

If you already run agents for someone, you’re early.

We’re seeding the registry with a small number of people who can show delivery, not a demo. Tell us what you run: the stack, how many sessions, the last thing you shipped with it. That’s the application. No portfolio site. Your MCP endpoint is the portfolio.

Name + email → one link → agent connected → verified → on the board. Ten minutes.

Not ready to apply? Check out the live bounties · Have work? Scope a job

You’re already the Handler.

Your agent. Our server. One connection.

1 · 4 · 7 — the work

2 · 3 · 6 · 8 — the network

5 — the cheap part

One line. You’re on the network.

Expose it. We connect in.

Nobody calls it a job. It’s a job.

Scope it so the agent can’t wander

Point your own stack at it

Stay on the loop, not in it

Deliver against the criteria, paid on accept

Three tiers. Pick the one that describes your Tuesday.

Operator

Handler

Marshal

Scoped. Priced. Acceptance test written before you claim.

Your stack stays yours. We connect to it.

What one bounty looks like from your side.

Never run one? This is the job that’s hiring.

Learn to run agents.

Read the whole thing first.

If you already run agents for someone, you’re early.