Which AI agents does this work with?

Anything with shell access — Claude Code, Codex, Cursor, Windsurf, Copilot CLI, Gemini CLI, or a loop you wrote yourself. There is nothing to install on our side and no integration to approve. Agents that can't run commands can call the same endpoints over HTTPS at /api/v1 with the same key.

What stops an agent from doing something expensive?

The binary, not the prompt. Destructive commands — all 11 of them — exit 2 without acting unless --yes was passed explicitly. Payout batches can't be created from the CLI at all, and the commands that would email your affiliates deliberately don't, so a retry loop can't spam anyone.

Where does the agent's API key live?

In its environment: TRACKREV_KEY always wins over anything saved, which is what automation wants. trackrev login instead stores a key at ~/.config/trackrev/config.json with mode 0600, verified against the API before it's written. Either way it travels only as a bearer token over HTTPS — the CLI refuses a non-https API base outside localhost.

Should the agent get its own key?

Yes. trackrev keys create --scope secret --label 'agent' prints the plaintext exactly once, to stdout only, so > agent.key captures just the key. Revoke it with trackrev keys revoke the moment you're done — it dies immediately and nothing else you run is disturbed.

Does the agent have to parse a pretty table?

It never sees one. Without a terminal the CLI emits raw tab-separated values — 2410.5, not 2,410.50 — and --json returns the API's own response body untouched, so the API reference doubles as the schema.

Does this need a paid plan?

Partly, and the split is exact: 43 of the 63 commands run on every plan including free — links, keys, webhooks, domains, folders, settings, trackrev me. The 20 that read analytics or move money need a paid plan and exit 3 on a free workspace, so 'wrong plan' never looks like 'no results'.

How does the agent learn the commands?

It runs trackrev --help. Every command, flag, plan requirement and confirmation rule is generated from one registry file — the same file this page's counts come from — so the help text can't drift from what the binary accepts.

TrackRev for AI agents

Your affiliate program, run by an AI agent.

Give Claude, Cursor, or any agent with a shell one secret key. It approves partners, answers revenue questions, builds launch links and reconciles commissions — and anything destructive stays locked until you type ‑‑yes.

Connect your agent See what it can run

$ npm i -g trackrev && export TRACKREV_KEY=lk_live_…

Any agent with a shell
One revocable key
Zero dependencies

your agent · trackrev

sample datashell access

Which channel actually made us money last week?

ran

$ trackrev channels --days 7 --ltv --json

4 channels · 2,143 clicks · $5,812.50 attributed

Newsletter: $1,890 from 310 clicks, against YouTube's $2,410 from 1,204 — about three times the revenue per click. Want me to weight the next campaign that way?

Commands

Locked behind --yes

Exit codes

Dependencies

Setup

Live in two minutes.

Nothing to install on our side, nothing to approve. Put the CLI where the agent runs, hand it a key, and it has your reach — same workspace, same plan limits, same audit trail.

TRACKREV_KEY in the environment — no config files, no OAuth
One labelled key per agent, revoked in one command
TRACKREV_API_URL points it at staging first

wherever your agent runs

bash

1# 1. The binary (Node 20+, zero dependencies)2npm install -g trackrev3 4# 2. A key of its own — the plaintext prints once, to stdout only5trackrev keys create --scope secret --label 'agent' > agent.key6 7# 3. Hand it over8export TRACKREV_KEY=$(cat agent.key)9 10# 4. Let it introduce itself11trackrev me --json

Open Settings → Developers to mint the first key by hand · package on npm

Claude CodeCodexCursorWindsurfCopilot CLIGemini CLIYour own loop

Listed because they can run a shell command — there is nothing of ours to install. No shell? The same surface is plain HTTPS at /api/v1, or MCP — npx -y trackrev-mcp.

Coverage

One key. The whole program.

Everything the dashboard does, as 63 commands — same limits, same attribution model, same audit trail.

Partners & programs

9 commands

Work the application queue, move affiliates between tiers, rewrite a program's rate and cookie window.

$ trackrev partners list --status pending --json
$ trackrev programs update prg_41 --rate 0.25

Commissions & payouts

5 commands

Read the ledger, record off-platform deals, void mistakes, settle payouts sent over PayPal or Wise.

$ trackrev commissions list --status pending --json
$ trackrev payouts list --status pending

Attribution & analytics

4 commands

Revenue per channel, per link, per click, per visitor.

$ trackrev channels --days 30 --ltv --json

Campaign links

7 commands

Launch links with expiry, passwords, QR — 500 at a time from CSV.

$ trackrev links create --url … --channel youtube

Revenue sync

5 commands

Connect Polar, Paddle, Lemon Squeezy & co. Pull charges on demand.

$ trackrev revenue sync --json

Visitors & orders

4 commands

Look anyone up by email; export any dataset as CSV.

$ trackrev visitors list --email @acme.com

Workspace setup

7 commands

The attribution model, the lookback window, the folders.

$ trackrev attribution set --model linear

Domains & pixels

7 commands

Branded short domains, DNS checks, retargeting pixels.

$ trackrev domains verify go.acme.com

Keys & webhooks

8 commands

Mint and revoke keys; point events at your stack.

$ trackrev keys create --scope secret

Email & branding

4 commands

Transactional emails on or off; partner-facing branding.

$ trackrev settings notify affiliate.approved --off

Orientation

3 commands

The first call: which workspace, which plan, which key.

$ trackrev me --json

Typing them yourself? The CLI page covers the same commands for humans.

In practice

Plain English in. Commands out.

No prompt pack to install — the agent reads trackrev --help and lands here.

“Where should next quarter's budget go?”

$ trackrev channels --days 90 --ltv --json

“Approve the affiliates who already sent traffic.”

$ trackrev partners list --status pending --json
$ trackrev partners approve ptr_8fd2 --program prg_41

“Why was this sale credited to email?”

$ trackrev visitors list --email sara@acme.io --json
$ trackrev journey 3f1b8c22-… --json

“One launch link per channel, expiring Dec 31.”

$ trackrev links create --url https://acme.com/launch --name Launch \
    --channel youtube --channel newsletter --expires 2026-12-31T23:59:59Z

“What do we owe partners right now?”

$ trackrev commissions list --status eligible --json

“Export the quarter's orders for finance.”

$ trackrev export --kind orders --days 90 --out q3-orders.csv

Guardrails

Destructive commands wait for a human.

Delete, revoke, ban, void, mark-paid — 11 of the 63 commands demand a confirmation an agent can't give. No terminal to answer on means exit 2 and nothing happens. Typing ‑‑yes stays your job.

No TTY and no ‑‑yes → refused, every time
Exit 2 is the same code as a typo — retry loops die here
Enforced by the binary, not by a system prompt

Mint a scoped key

agent, no tty

$ trackrev links delete lnk_3b91

Delete lnk_3b91? This cannot be undone.

Refusing to run a destructive command non-interactively. Pass --yes.

exit 2 · nothing was deleted

after you say yes

$ trackrev links delete lnk_3b91 --yes

Deleted lnk_3b91

Four exit codes

0 ok · 1 failed · 2 misused · 3 wrong plan. Agents branch instead of guessing.

stdout is only data

TSV when piped, --json verbatim. Warnings and prompts go to stderr.

Keys you can yank

Scoped, labelled, last-use audited. Revoked in one command, dead instantly.

No payout batches

Fees and payout floors are computed in one place — the dashboard.

No surprise emails

CLI approvals don't email affiliates, so a retry loop can't spam them.

No Stripe access

Your restricted key lives on the workspace, out of the CLI's reach.

The brief

Brief it once.

Six rules cover it: where the key lives, how to discover commands, what each exit code means, and where the line is. Paste it into AGENTS.md, CLAUDE.md, or a system prompt — the file name changes, the content doesn't.

--help over guessed flags
Exit 2 means re-read, not retry
--yes, rates and payouts stay human

AGENTS.md

markdown

1# TrackRev2 3Our affiliate program and link attribution live in4TrackRev. Use the `trackrev` CLI (npm i -g trackrev).5The key is in TRACKREV_KEY — never print it.6 7- Run `trackrev --help` for the command list and8  `trackrev <command> --help` for one command's9  flags. Never guess a flag.10- Always pass --json. stdout is data; warnings and11  captions go to stderr.12- Exit codes: 0 ok · 1 failure · 2 you used it wrong13  · 3 needs a paid plan. On 2, re-read --help rather14  than retrying. On 3, stop and tell me.15- Open with `trackrev me` to confirm the workspace.16- Destructive commands (delete, revoke, ban, void,17  mark-paid) refuse to run without --yes. Never add18  --yes yourself. Show me the command and wait.19- Money and terms are mine: don't change commission20  rates, attribution settings or program status21  without asking first.

Unattended

Runs while you sleep.

Cron it, CI it, or let the agent schedule itself — exit codes make failures loud.

Morning numbers, ready to summarise

bash

trackrev channels --days 1 --json > /tmp/yesterday.json
trackrev commissions list --status pending --json >> /tmp/yesterday.json

Flag waiting affiliates once a day

bash

trackrev partners list --status pending --json \
  | jq -e '.partners | length > 0' && echo 'affiliates waiting'

Gate the deploy on DNS

bash

trackrev domains verify go.acme.com   # exits 1 while pending

Give it a key, then take it back

bash

trackrev keys create --scope secret --label 'agent' > agent.key
trackrev keys revoke key_3f19 --yes

FAQs

Questions, answered.

What teams ask before giving a model access to the system that decides who gets paid.

Questions8

Do you have an MCP server?

TrackRev

Yes — npx -y trackrev-mcp, 61 tools over stdio, nothing to host. It is generated from the same registry the CLI is, so the two can't drift: every command is a tool, flags become typed parameters, and a parameter that isn't real is rejected by name instead of being ignored. Destructive tools take a confirm that must be true, and --read-only drops everything that can change data. If you'd rather not install anything, the same surface is the CLI or plain HTTPS at /api/v1.

Go deeper

Give it a key. Watch it work.

Mint a scoped key, drop it in your agent's environment, and start read-only — trackrev channels --days 7 --json is a fair first task. Revoke the key whenever you like.

Start free Read the CLI docs

MIT licensed · Zero dependencies · Node 20+

Your affiliate program, run by an AI agent.

Live in two minutes.

One key. The whole program.

Partners & programs

Commissions & payouts

Attribution & analytics

Campaign links

Revenue sync

Visitors & orders

Workspace setup

Domains & pixels

Keys & webhooks

Email & branding

Orientation

Plain English in. Commands out.

Destructive commands wait for a human.

Four exit codes

stdout is only data

Keys you can yank

No payout batches

No surprise emails

No Stripe access

Brief it once.

Runs while you sleep.

Questions, answered.

Pairs well with

Channel Analytics

Visitor Journey Timeline

Click Analytics Breakdown

Give it a key. Watch it work.

Your affiliate program, run by an AI agent.

Connect an agent

Live in two minutes.

One key. The whole program.

Partners & programs

Commissions & payouts

Attribution & analytics

Campaign links

Revenue sync

Visitors & orders

Workspace setup

Domains & pixels

Keys & webhooks

Email & branding

Orientation

Plain English in. Commands out.

Destructive commands wait for a human.

Four exit codes

stdout is only data

Keys you can yank

No payout batches

No surprise emails

No Stripe access

Brief it once.

Runs while you sleep.

Questions, answered.

01Do you have an MCP server?

02Which AI agents does this work with?

03What stops an agent from doing something expensive?

04Where does the agent's API key live?

05Should the agent get its own key?

06Does the agent have to parse a pretty table?

07Does this need a paid plan?

08How does the agent learn the commands?

Pairs well with

Channel Analytics

Visitor Journey Timeline

Click Analytics Breakdown

Give it a key. Watch it work.