A builder's wishlist: ten under-served agents the Swarms Marketplace is missing, each with a spec, the MCP servers it should use, the swarm architecture that fits, and a pricing model. Build one and launch it in a competition.
The Swarms Marketplace has more agents than anyone could read in a week, and a large share of them do the same few things. Meanwhile 25 official MCP servers from Google, PayPal, Neon, PagerDuty, Perplexity and others landed in one drop on October 20. The servers are there. The agents that put them to work mostly are not.
This is a wishlist: ten agents we think should exist and do not, each with a short spec, the MCP servers it should use, and the Swarms API architecture that fits. These are the listings we feature, and the kind of work our competitions exist to put in front of people. Build one and we want to see it.
The Explorer gives the live numbers. When that post was written, on October 17, 2026, the marketplace listed 6,400 products from 808 sellers: 5,576 agents, 773 prompts and 46 tools. 214 products were paid, 996 were tokenized, and 24 agents exposed an MCP server.
Agents outnumber tools by more than a hundred to one, so the agent type is where the crowd is. And only 24 agents serve an MCP server, so the link between the marketplace's agents and the systems people run their businesses on is thin. The Explorer counts by type, so this part is our impression from browsing the registry: plenty of general assistants, personas and trading-adjacent agents, and very few that log into a specific system of record and do one bounded job.
The infrastructure for that second kind arrived in bulk. mcp_depot published 25 vendor-hosted servers on October 20, bringing its catalog to 55, across databases, payments, search, incident management, project management, meetings, content and infrastructure. Every agent below uses at least one.
Each one has a buyer with a budget and an output that can be checked. The Auto Agent Builder turns any problem statement below into a first-draft roster in one call.
Problem. Every team has a migration it keeps postponing: a framework major version, a database move, a CI swap. The hard part is the plan that names every affected file, orders the steps so the system keeps working, and says how to roll back. Nobody wants to write it, so the migration slips.
PlannerWorkerSwarm. The steps cannot be listed in advance, and the built-in judge keeps re-planning until the plan covers the goal.Demo: point it at a public repo and publish the plan, including the step where a dry run on a Neon branch failed.
Pricing: one-time paid listing by card. A plan that saves two days of a senior engineer's time is worth well above the usual agent price.
Problem. A business that invoices through PayPal and takes counter payments through Square ends the month with three records that disagree. Someone matches rows by hand. The matches are easy; the exceptions are the job.
SequentialWorkflow. Extract, normalize, match, explain. Each step needs the one before it, and one turn per agent doubles as an audit trail.Demo: a synthetic month of 200 transactions with a dozen planted exceptions, shown next to the agent's list.
Pricing: one-time paid listing by card. The fiat rail exists for this buyer: someone who expenses a tool and never thinks about a wallet.
Problem. The incident ends at 3 a.m. and the writeup is due at the 10 a.m. standup. What exists is a PagerDuty timeline, a pile of chat and a few support tickets, and turning that into a report lands on the most tired person in the room.
MixtureOfAgents. Timeline, impact and follow-up agents read the same evidence in parallel, and the aggregator writes the one report that gets pasted into standup.Demo: a sandbox incident turned into a two-page report, with follow-ups visible on a board.
Pricing: paid by card, or tokenized with Vault Mode so the on-call rotation holds the token and gets access (how Vault Mode works).
Problem. Analytics queries get written once and run forever. A query that scans a whole ClickHouse table every five minutes, or a BigQuery job that reads a terabyte to return ten rows, is a quiet bill nobody reviews.
ConcurrentWorkflow. The reviews are independent and the buyer wants them separately, so there is nothing to merge and no reason to run them in a line.Demo: ten public-dataset queries before and after, with bytes scanned reported by the database itself.
Pricing: free listing with a tokenized launch. Developers will run it on every query, which suits a creator share of every trade better than a one-time price.
Problem. Most research agents return links with a paragraph under each. A decision maker needs a position: what the evidence says, where sources disagree, which claims were checked against a primary source, and what is unknown.
HeavySwarm. Research, analysis, alternatives and verification run as separate workers, and the synthesis agent writes the brief. The most expensive shape here, and the right one.Demo: one contested technical question where the top results disagree, with the disagreements table doing the work.
Pricing: one-time paid listing by card, with a free companion prompt showing the output format.
Problem. Every prompt on the marketplace now carries a Security Score that catches malicious instructions. Nothing checks a prompt against a company's own rules: regulated language, claims it may not make, data it must not request, disclosure requirements. Finance, healthcare and legal teams do this by hand today.
MajorityVoting. Each rule is a yes or no question, independent voters cut single-model variance, and the vote count is the evidence. CouncilAsAJudge's six fixed dimensions are the wrong rubric here.Demo: twenty prompts from the public registry run against one published policy, with the results table and redlines.
Pricing: one-time paid listing by card, for compliance and platform teams.
Problem. A product meeting produces eight decisions and five action items. Two reach the tracker. Recording tools already summarize; the missing step turns a summary into tickets that fit the board without duplicating work already filed.
SequentialWorkflow. Extract decisions, draft, dedupe against the board, file. The order is fixed and every step depends on the last.Demo: a 40-minute meeting, the board before and after, and the duplicate it refused to create.
Pricing: one-time paid listing by card at the low end of the agent range, so a team lead can buy it without approval.
Problem. Translating a site is the easy half. Checking it is where the cost sits: strings overflow buttons, product names get translated when they should not, one term gets three renderings. Human QA for ten languages costs more than the translation did.
GroupChat. A linguist, a terminology agent and a layout-constraints agent review one page together, and a length constraint changes the linguist's fix. Keep it to three or four agents.Demo: one public Webflow site in two languages, with the issues list and at least one protected term it caught.
Pricing: one-time paid listing, or a tokenized launch quoted in USDC. The USDC Launch Championship post explains why dollar pricing suits buyers with budgets.
Problem. Pipelines fail quietly. A table stops updating, row counts halve, a column goes null, and nobody notices until a dashboard looks wrong. The expensive failures are the ones where every job reported success.
MultiAgentRouter. Each anomaly type has a specialist and most runs trip only one, so agents the router skips never run.Demo: a public dataset with one planted freshness failure and one planted volume drop, and the two different diagnoses.
Pricing: free listing with a tokenized launch. The teams that adopt it are the people who will hold the token.
Problem. Nonprofits, labs and small vendors answer 60-page funding calls that ask the same twenty questions in new words every time. The answers already exist in past proposals in a shared folder, and adapting them is a week of a grant writer's time per submission.
HierarchicalSwarm. The director reads the call, splits it into sections, hands each to a writer with the relevant past material, and integrates the result. Splitting a 60-page call is itself a judgment call, which is what a director is for.Demo: a real public RFP answered from a synthetic set of past proposals, with the requirements matrix shown in full.
Pricing: one-time paid listing by card at the top of the agent range. One funded proposal pays for it.
There is no separate submission form. The path is the one every competition post describes.
publish_to_marketplace=True and tags set (example here).The marketplace has 5,576 agents and 25 new official MCP servers waiting for them. The next few hundred agents decide whether the catalog reads as a pile of demos or a set of tools people pay for. Build one of these ten.
A one-hour walkthrough for new visitors to swarms.world: find an agent with Autonomous Filter Search, read its listing, try it in Swarms Chat, pull it into your code through the Marketplace MCP and the Swarms API, and buy a paid listing by card, SOL, or USDC.
A hands-on tutorial: wrap the keyless Open-Meteo weather API in a Python MCP server with FastMCP over Streamable HTTP, test it, add a bearer-token check, ship it in Docker, and publish it on the Swarms Marketplace MCP Portal so agents can find and call it.
A migration guide for developers sitting on SKILL.md folders, .cursorrules files, or Custom GPT configurations: what maps to a prompt, agent, or tool listing on the Swarms Marketplace, what to strip first, how to import from private GitHub repositories, and what is worth charging for.