Sequential
Each agent hands its output to the next. Cheapest to reason about, slowest end to end.
Hivecraft/Workspace/research-desk
Workspace · a working design, runs in your browser
Draw the swarm. Price it before it runs. Sell it by the call.
Hivecraft is a working canvas for multi-agent workflows. Drag nodes, wire them up, run a timed simulation, then take the graph straight into the cost and revenue models below.
CA
Add node
Drag a node by its header · drag from a right-hand port to connect · click a wire, then Delete.
Empty canvas. Pick a template or add nodes from the left.
0 nodes
0 wires
critical path —
cost / run —
—
swarm id —
Simulation log model timings, nothing is called
Anatomy of a run
One request, seven cells, every token on the meter.
- t+0.0sRequest in
A question enters the swarm and a search tool fetches sources. Almost nothing is spent yet.
- fan-outThree analysts at once
The same sources go to three agents in parallel. Parallel work buys speed, not savings.
- tokensThe meter runs
Each analyst reads and writes. Cost follows tokens; time follows the slowest branch.
- mergeReconcile and settle
A large model merges the three briefs, and one Solana transfer settles the run.
- Tokens
- 0
- Model cost
- $0.0000
- Elapsed
- 0.0 s
Model · sample list prices from the Run costs table, no cache. Nothing is called.
Patterns
Five shapes cover most swarms worth paying for.
Each shape trades money for something: parallel workers buy speed, routers buy focus, critics buy quality. Load any of them onto the canvas and the cost model updates with it.
Fan-out, merge
Parallel workers attack one question from different angles; a merge step reconciles them. Fast, token-hungry.
Router
A cheap classifier sends each request to one specialist. Expected cost drops with every branch you add.
Critic loop
A reviewer sends drafts back until they pass. Quality rises; so does the bill — set the iteration cap honestly.
Manager + workers
A planning agent splits the task and writes the brief for each worker. Best when tasks are uneven.
Marketplace
Agents, tools and prompts you can drop onto the canvas.
The listings below are sample entries written to show how the catalogue works. Filters, search, details and “add to canvas” all function.
Run costs
What one run of this swarm costs, node by node.
The estimator reads the graph on the canvas. Token prices are sample list prices you can edit. Execution probability follows the graph: a router's branches each run a share of the time.
Assumptions
Reading priority fees…
Model price table sample, USD per 1M tokens
| Tier | Input | Output | Tok/s |
|---|
Per run
—Per day
—Per 30 days
—Seller revenue
Price it, then see what is left after the platform, the models and the refunds.
Subscription mode: each subscriber is assumed to make the number of calls below per month, which drives model cost.
- Gross, month 1
- —
- Platform fee
- —
- Refunds
- —
- Model cost
- —
- Net payout, month 1
- —
- In SOL at live price
- —
- Margin per unit
- —
- 12-month net
- —
Bars: monthly net (left scale) · dashed line: cumulative net (right scale). Model — projections, not earnings data; the platform fee is a proposal.
Manifest
A listing is a graph plus a price, fingerprinted.
When a swarm is published, buyers see exactly which nodes, models and tools it runs. The manifest below is generated live from your canvas and seller settings; its id is the SHA-256 of the canonical graph, so any change to the swarm changes the id.
Agent pulse
The on-chain agent economy, read live.
A small basket of Solana AI-agent tokens from DexScreener, refreshed every minute. It is market data for context — not a ranking, recommendation or endorsement.
Basket 24 h volume
—Volume-weighted 24 h
—Buys / sells 24 h
—Pool liquidity
—| Token | Price | 24 h | Volume 24 h | Buy pressure | Liquidity |
|---|---|---|---|---|---|
Source: DexScreener tokens API · deepest pool per token
FAQ
Straight answers.
Does the canvas actually run agents?
No. “Simulate run” walks the graph in dependency order and times each node with the model's latency figures. No model, tool or API is called, and nothing leaves your browser.
Where do the prices in the cost model come from?
They are sample list prices per million tokens for four generic tiers, chosen to be in a realistic range. Every value is editable; replace them with your provider's current rates. The Solana settlement cost uses the live priority-fee reading and live SOL price.
Are the marketplace listings real?
No. They are sample entries that show how listings, filters and detail sheets work. Hivecraft is a working design, not a live marketplace, and no listing can be bought here.
What is the swarm id?
A SHA-256 hash of the graph's canonical form: node kinds, labels, models, token budgets and wires, without screen positions. Moving a node keeps the id; changing what the swarm does changes it.
Can I take my graph elsewhere?
Yes. Export JSON or YAML from the toolbar. The format is plain and documented in the manifest panel. JSON can be imported back.
How is expected cost handled for routers?
Each outgoing branch of a router is assumed to run an equal share of the time. A node reached by several paths is capped at running once per run.
Is anything saved?
Your canvas is saved in this browser's local storage so it survives a reload. There are no accounts, cookies or analytics.
Is there a token?
If one is listed, its address appears at the top and in the footer. Nothing on this site mints, trades, holds or signs anything.