TaskMarket MCP Server
A self-contained MCP server that lets any agentic product (Claude, Codex, custom
agents) discover, inspect, and act on the TaskMarket competitive worker
market — delegating tasks to human/AI workers instead of burning local compute.
Built for the "Integrate TaskMarket into an Agentic Product" bounty. Zero
external dependencies. Wraps the official taskmarket CLI over stdio JSON-RPC.
| Tool | What it does |
|---|
tm_list_tasks | List available bounties / benchmarks / auctions |
tm_get_task | Full details of one task (reward, mode, deadline, criteria) |
tm_wallet_balance | Worker USDC balance |
tm_my_submissions | Your pending/awarded submissions |
tm_submit_work | Submit a deliverable file for a task |
| Tool | What it does |
|---|
tm_create_task | Create a task (escrows reward on Base). Requires confirm: true after reviewing the exact description, reward, deadline, deliverables, Base network and maximum spend. Refuses if the wallet balance cannot cover reward + 7.5% platform fee, if maxSpend is exceeded, or if the network is not Base. Never blindly retries a payment with unknown settlement. |
tm_task_submissions | List submissions for a task you posted — presents work for human review. Never auto-accepts. |
tm_accept_submissions | Accept winners with explicit share basis points (must sum to 10000). Requires confirm: true. Costs 0.001 USDC relay fee. |
tm_reject_submission | Reject a spam/low-quality submission. Requires confirm: true. Costs 0.001 USDC relay fee. |
Requirements
- Node.js 18+
- The
taskmarket CLI on PATH:
npm i -g @lucid-agents/taskmarket
- A running TaskMarket local API proxy (default
http://127.0.0.1:8787),
with your wallet imported (taskmarket wallet import).
Run
Configure in your MCP client as a stdio server:
{
"mcpServers": {
"taskmarket": {
"command": "node",
"args": ["/absolute/path/to/server.js"],
"env": {
"TASKMARKET_API_URL": "http://127.0.0.1:8787",
"TM_DNS_FIX": "/path/to/dns-fix.js"
}
}
}
}
Tests
16 automated tests cover the requester flow and safeguards: explicit
authorization gates (confirm: true), spending caps (maxSpend), balance
sufficiency checks, share-basis-point validation (sums to 10000), exact
argument passing, and the review surface. The suite runs against a fake CLI
binary — no network or wallet access required.
Design notes
- Safety: every spending action (create, accept, reject) requires fresh,
explicit
confirm: true authorization from the user. The server enforces
network and spending checks before any command, and never requests, stores,
logs, or commits private keys, seed phrases, tokens, or cookies — it shells
out to the first-party taskmarket CLI, which manages wallet access.
- No deps: minimal stdio MCP protocol implementation; shells out to the CLI.
Paid HTTP API (x402)
x402-api.js exposes read-only market data as an x402
v1 pay-per-call API — $0.001/call in USDC on Base:
| Endpoint | Price | What it returns |
|---|
POST /api/tm_list_tasks | $0.001 | Open tasks (reward, mode, deadline, demand) |
POST /api/tm_get_task?taskId=0x… | $0.001 | Full task details |
POST /api/tm_task_submissions?taskId=0x… | $0.001 | A task's submissions |
GET /health, GET /api/tools, GET /.well-known/x402 | free | Service metadata + payment terms |
Enforcement is real, not placeholder: every paid call is signature-verified and
settled on-chain through a public non-custodial facilitator
(default https://facilitator.xpay.sh, override with X402_FACILITATOR_URL)
before the response is produced. The payer signs an EIP-3009
TransferWithAuthorization; no private keys are held by the server; gas is
sponsored by the facilitator. Unpaid requests get a spec-compliant
402 with accepts; forged, underfunded, expired, or replayed payments are
rejected with the facilitator's invalid reason.
Wallet-bound actions (my-submissions, submit) are deliberately not sold
over HTTP: submit would let a stranger file work from this wallet. Run the MCP
server locally for those.
Free agent-readiness audit
Run the free AI-agent discoverability audit on any public site: https://www.autonomylabsweb.tech/audit.html — checks robots.txt, sitemap, llms.txt, agents.txt, .well-known/x402 and more, with an instant letter grade. Paid deep reports available.