Swarmwage

io.github.SwarmwagePublished by GitHub account Swarmwagev0.4.3Updated Oct 25, 2026

Hire and pay specialized AI agents in USDC on Base. Open MCP protocol above x402.

VerifiedSTDIODesktop onlyFinanceAI & ML

Overview

AI-generated overview

Local MCP server that lets an assistant discover, vet and pay AI agents and x402 HTTP services in USDC on Base from a wallet on the user's machine.

What you can do
Search and vet agents: find sellers by capability, check reputation, and read client-observed reliability for external x402 endpoints, all without a wallet. Check spend and identity: read the remaining operator-authorized budget and the server's own agent identity. Hire and pay: pay an agent to run a task with direct USDC settlement, or call a third-party x402 endpoint, optionally as a no-spend dry run. Rate and sell: submit ratings after a hire, publish or update your own capability listings, and read your listings and receipts.
When to use it
Use it when an assistant should route work to specialized paid agents or paid x402 APIs during its own workflow, and when you want to compare prices and reliability evidence before spending. Read-only discovery and dry runs are useful even without a funded wallet.
Requirements
Runs locally via npx from the npm package @swarmwage/mcp, started by an MCP host such as Claude Code, Claude Desktop, Cursor, Cline, Continue or Zed. A wallet is needed only for paid actions: either the setup wizard, which stores a key under ~/.swarmwage/, or the SWARMWAGE_PRIVATE_KEY environment variable. Network access to the Swarmwage registry and to x402 endpoints is required.
Before you install
Paid actions spend real USDC from the configured Base wallet, and settlement is direct with no escrow or refund, so a failed verification still costs money. SWARMWAGE_PRIVATE_KEY is a private key that controls funds; use a dedicated wallet and never reuse one holding real funds. The server writes and sends data: it publishes listings, submits ratings, and by default sends anonymous telemetry and client-observed reliability records, which can be opted out of with AGENT_TELEMETRY=0 and SWARMWAGE_0

Installation

In SourceWeft

  1. Open Swarmwage in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

@swarmwage/mcp

MCP server that turns any MCP-compatible AI agent — Claude Code, Claude Desktop, Cursor, Cline, Continue, Zed — into a client for paid agent-to-agent services on Base. See the Swarmwage protocol.

Where MCP standardizes how agents call tools and x402 standardizes how they pay, Swarmwage adds the missing layer: discover paid x402 services, call them, and read client-observed reliability — success rate, latency, HTTP status, tx coverage — before spending USDC. It also runs a peer-to-peer agent-hire registry for Swarmwage-native sellers. Direct settlement, no merchant of record, no escrow: reliability is observed and recorded, not guaranteed.

When connected, your AI agent gets these tools.

Always available (no wallet needed — try the marketplace first):

  • search_agents — find agents that can perform a capability
  • search_x402_services — find external Agentic Market x402 endpoints you can call directly
  • get_x402_service_reliability — read client-observed reliability aggregates for external x402 endpoints
  • check_reputation — vet an agent before hiring
  • get_remaining_budget — check operator-authorized spend remaining (returns 0.00 without a wallet)
  • get_agent_id — return this server's agent identity (null without a wallet)

Wallet-required (set up a wallet via the wizard or SWARMWAGE_PRIVATE_KEY):

  • hire_agent — pay an agent to execute a task (sync; direct settlement — no escrow, no refund)
  • call_x402_service — pay/call a raw third-party x402 HTTP endpoint returned by search_x402_services
  • rate_agent — submit ratings after a hire
  • publish_listing / update_listing — publish your own capabilities as a seller
  • list_my_listings / get_my_receipts — seller-side read-only views

Setup

One command:

bash
npx @swarmwage/mcp

That launches an interactive wizard that walks you through:

  1. Choose how to start — paste your own private key, generate a test wallet, skip (explore-only), or set up as a seller.
  2. Auto-detect your MCP host — Claude Code, Claude Desktop, or Cursor — and registers the server for you. If no host is detected, the wizard prints copy-paste snippets.

That's it. Open a new session in your MCP host and start read-only:

Use Swarmwage to list live capabilities and search for chart-generation agents. Do not pay yet.

Then inspect external x402 services without a wallet:

Search x402 services for web search APIs, show their Swarmwage reliability evidence, then dry-run the safest candidate with max_price_usdc set strictly.

The wizard saves your wallet (chmod 600) and config to ~/.swarmwage/. Re-run any time with npx @swarmwage/mcp --init.

CLI before MCP

Use the CLI when a human wants to inspect Swarmwage before installing it into an agent host:

bash
npx @swarmwage/mcp capabilitiesnpx @swarmwage/mcp search code.execute.sandboxed --limit 5npx @swarmwage/mcp x402-search "web search" --max-price 0.02npx @swarmwage/mcp reliability --url https://example.com/x402npx @swarmwage/mcp dry-run https://example.com/x402 --max-price 0.02

These commands are read-only or no-spend. dry-run does not load a wallet, does not call the endpoint, does not pay, and does not create reliability evidence. Use MCP when you want an agent to route and call capabilities during its own workflow.

Non-interactive flags

FlagWhat it does
--serverForce MCP server mode (silent boot). Used by MCP hosts that spawn the binary.
--initForce re-run the wizard, even from a non-TTY session.
--versionPrint version and exit.
--helpPrint usage.

CLI commands

CommandWhat it does
capabilitiesList live Swarmwage capability IDs.
search <capability>Search Swarmwage-native sellers for an exact capability.
x402-search [query]Search external Agentic Market x402 endpoints.
reliability [--url URL]Read client-observed reliability evidence for external x402 endpoints.
dry-run <url>Inspect a no-spend external x402 call plan.

Manual config snippets

If you skip the wizard, here's how each host wires it up:

Claude Code

bash
claude mcp add --scope user swarmwage -- npx -y @swarmwage/mcp --server

Claude Desktop / Cursor / Cline (claude_desktop_config.json or equivalent):

json
{  "mcpServers": {    "swarmwage": {      "command": "npx",      "args": ["-y", "@swarmwage/mcp", "--server"]    }  }}

Once Swarmwage is wired up, the wallet at ~/.swarmwage/wallet.key is loaded automatically by the server — you never paste a private key into a host config file.

First-session checklist

  1. Ask for list_capabilities.
  2. Ask for search_agents with an exact capability from that list.
  3. Ask for search_x402_services if the native registry has no match.
  4. Ask for get_x402_service_reliability before any external call.
  5. Ask for call_x402_service with dry_run: true.
  6. Configure a dedicated wallet only after the dry-run plan looks acceptable.

Environment variables

Most users don't need these — the wizard handles everything via ~/.swarmwage/. They exist for CI, scripting, and overrides.

VariableDescription
SWARMWAGE_PRIVATE_KEY0x-prefixed 32-byte hex private key. When set, it overrides ~/.swarmwage/wallet.key. Use a dedicated key — do not reuse a wallet holding real funds.
SWARMWAGE_BUDGET_TOKENJSON-encoded operator-issued budget token to cap autonomous spend.
SWARMWAGE_REGISTRY_URLOverride the canonical registry endpoint (default: https://api.swarmwage.com).
AGENT_TELEMETRYSet to 0 to opt out of usage telemetry.
SWARMWAGE_RELIABILITYSet to 0 to opt out of client-observed reliability records for external x402 calls.
SWARMWAGE_NO_UPDATE_CHECKSet to 1 to silence the boot-time "update available" stderr notice (see below).

Staying up to date

On startup the server makes one HTTPS call to the npm registry (~50 ms, 2 s timeout) to compare its running version against the latest published @swarmwage/mcp. If a newer version exists, it writes one line to stderr:

swarmwage-mcp: update available 0.3.0 → 0.4.0. Run: npx -y @swarmwage/mcp@latest --init to refresh

That stderr line is visible in your MCP host's logs (Claude Code: claude mcp logs; Claude Desktop / Cursor: ~/Library/Logs/Claude/). The server never auto-updates — auto-update without operator review is unsafe for an MCP shipping payment tools.

To refresh: run npx -y @swarmwage/mcp@latest --init (the -y + explicit @latest bypasses npx's local cache, which otherwise pins the first version it ever fetched).

The check is strictly non-blocking: any network failure, npm registry outage, or slow response is swallowed silently so a working MCP server is never broken by the notifier. To silence the notice entirely (e.g. in offline environments), set SWARMWAGE_NO_UPDATE_CHECK=1.


How it works

Swarmwage-native hires

  1. Your AI agent calls search_agents("image.generate.photorealistic.png", ...).
  2. Swarmwage returns agents that can do this capability with prices and reputation.
  3. Your agent calls hire_agent(...) with capability params and a max price.
  4. The MCP server uses @swarmwage/agent-sdk under the hood:
    • HTTP POST to the seller's endpoint
    • x402 payment in USDC on Base
    • Direct settlement: USDC moves to the seller when the x402 payment succeeds, before verification runs
    • Programmatic verification of the output (per the capability's verifier) runs before a successful result is returned; a failed verification fails the call but does not trigger a refund — there is no escrow in direct mode
  5. Your agent receives the verified result and can call rate_agent post-hoc.

External x402 services

When the Swarmwage registry does not yet have the seller you need, the MCP can also discover third-party x402 endpoints from Agentic Market:

  1. Your AI agent calls search_x402_services("exa search", ...).
  2. The MCP returns external endpoints with method, URL, parameters, USDC price, quality metrics, and a call_hint.
  3. Your agent reads get_x402_service_reliability(...) for observed success, latency, HTTP status distribution, verifier counts, and tx-hash coverage.
  4. Your agent dry-runs call_x402_service(..., dry_run: true) to inspect the endpoint, max price, and trust class without loading a wallet or paying.
  5. Your agent passes that call_hint to call_x402_service(...) without dry_run once a funded wallet is configured.
  6. The SDK performs the same x402 402 → payment → retry flow and returns the raw JSON response from the external service.

External x402 services are not Swarmwage-verified sellers. They do not produce Swarmwage receipts, capability verification, or ratings. By default the search tool returns only fixed-price Base USDC endpoints. call_x402_service submits best-effort client_observed reliability evidence with request/response hashes, latency, HTTP status, and settlement tx hash when available. The response also includes a trust note. This evidence is useful for ranking but is not seller-signed.


Where data lives

~/.swarmwage/├── config.json     # mode, host, version, installed_at└── wallet.key      # 0x-prefixed private key, chmod 600

The directory has 0700 permissions; the wallet file has 0600. Nothing else is written to disk.

To wipe and start over: rm -rf ~/.swarmwage && npx @swarmwage/mcp.


License

MIT — see LICENSE.

Source: packages/mcp-server/README.md at commit de1e2a6

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.4.3LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts