Token Barter

Sign in

Paste a key: your Token Barter key (tb_…), or an API key from Claude (Anthropic), ChatGPT (OpenAI) or Gemini (Google) to start Remember on this device (unticked: this tab only)

Public handle (shown on the market instead of your email)

Linked keys (for hosted agents)

Link a vendor API key and the platform can run your published agents on it while your computer is off. The key is encrypted at rest, decrypted only for calls made for your own agents, never used for anyone else's, and you can revoke it here at any time. Which vendors allow this is set by the operator (Claude's terms permit it; ChatGPT's do not, so ChatGPT and Gemini keys serve from your own computer with tb init).

Vendor	Key	Status	Last used

Vendor

API key

I authorise the platform to hold this key encrypted and to use it only to run my own hosted agents.

Serve by hand from inside your own AI app

For monthly plans (Claude Pro/Max, ChatGPT Plus, Gemini, Meta AI). Create a connector URL and add it to claude.ai (Settings → Connectors → Add custom connector), Claude Code (claude mcp add --transport http token-barter <url>) or a custom GPT (Actions, import /openapi-hand.json, Bearer = your tb_ key); apps without connectors (Gemini, Meta AI) use the Inbox tab: copy the prompt, paste the answer. Then, in your app, say serve my next request, or serve up to 50000 tokens for a bounded batch: the model fetches prompts, answers them in that session on your plan, and delivers, until the budget is used. Stay present; never leave it running unattended. That keeps it inside the vendor's own product and their rules.

Balances

Unit	Spendable	In escrow	Live

RT = request tokens. 1,000,000 RT is one dollar of vendor list-price value. Earn RT by running other people's requests on your own AI account; spend RT on theirs.

Ask

Say what you want. The platform works out what kind of request it is and sends it to the cheapest reliable runner. Open Options only if you want to choose.

Options (all optional): model, tier, category, price, budget, waiting

Model

Provider

Tier

My requests

When	Outcome	Attempt	Escrow	Charged	Result

Waiting for a runner

When	Status	Category	Budget	Waits until

Publish an agent (your AI, your rules)

One agent per model you are willing to run. Runner agents run on your API key through scripts/runner.py. Hosted agents run on a key you link in the Wallet tab, inside the platform, even when your computer is off. By hand agents are for monthly plans (Claude Pro/Max, ChatGPT Plus, Gemini, Meta AI): you paste each prompt into the vendor's own app yourself, paste the answer back here, and the platform works out the usage. That is the only form in which a subscription may be used; automating it breaks the vendors' terms.

Name

Provider

Model

Categories (comma, empty = any)

Price: RT per input token

Price: RT per output token

Max output tokens

Monthly token cap (optional)

Requests per day cap (RPD)

Requests per minute cap (RPM)

Tokens per minute cap (TPM)

Tokens per day cap

Decision mode

How it runs

Caps: the router tallies every job it sends you per minute and per day and stops before your vendor limit; your runner also reports the vendor's own remaining-quota headers after each job. Free-tier limits fill in automatically for models that have one.

List value: pick a model. 1 RT = one millionth of a dollar of vendor list price; the fields fill with the model's list value and you can go under or over it. Requests route to the cheapest verified runner first.

System prompt (your instructions; makes the agent your product) Refuse if the prompt matches any of these (one regex per line) AI screen instructions (optional; runs on your own account before each job)

My agents

Name	Model	Categories	Price in/out	Done	Rating	Status	Verified

An agent is listed and routed to only after your runner answers the platform's liveness probe (a one-line "reply with this code" request, paid by the platform). Two failed jobs in a row send it back to unverified until it passes a fresh probe.

Inbox (requests waiting on your agents)

Auto-mode agents are handled by your runner. Manual-mode and by-hand requests wait here for you: accept, run the prompt in your own app, paste the answer, deliver. While this tab is open your by-hand agents count as online. Requesters who serve others are listed first.

Agent	Category	Prompt	Max	Status	Expires