Most MCP servers expose a fixed tool list: here are 10 actions you can call. That is safe and predictable — and it quietly kills negotiation.
An intelligent MCP client can reason about what it needs. A human bakery customer can say: “like loaf B, but less sugar.” A fixed MCP catalog usually answers with silence: no matching tool → agent leaves. The seller never learns the demand; the buyer wastes a discovery hop.
AgentShare (agentshare.dev) is an agent-paid API (dual-auth: API key or x402) with free discovery and MCP Streamable HTTP. We still ship a fixed catalog — but we added a deliberate gap-closer:
When existing tools are insufficient, agents can submit a structured capability request:
capability — what is missing desired_input / desired_output — contract sketch why_insufficient — why current tools fail budget_usd — willingness to pay per call if we ship it
Optional: nearest_existing_tool, contact_hint (for our admin web inbox — we do not auto-email).
Safety gates (why this is not “AI writes new code”)
Engineers freeze tool lists for good reasons: security and predictability. Letting a client invent executable tools at runtime is a non-starter.
So request_capability is intake + human review, not codegen:
- Auth — API key or x402
- Rate limits — per key / per IP
- Mandatory schema — vague wishes are rejected
- Paid stake — catalog price ~$0.05 USDC (spam filter)
- Admin web inbox on agentshare.dev — fulfill, plan, or decline
Closest existing tools are still preferred; this channel captures paid demand signal when the menu is wrong.
Try it
- MCP:
https://agentshare.dev/mcp
- REST:
POST https://agentshare.dev/api/v1/agent/capabilities/request
- Discovery:
https://agentshare.dev/agent.json · https://agentshare.dev/.well-known/x402
- Public face /
llms.txt: github.com/anhmtk/agentshare-mcp
Fixed tools stay the default path. Negotiation is now a first-class, gated tool — the bakery counter, not the unlocked kitchen.