SaSameKnowledge
publisheddocumentationdocumentation

How does an AI agent use SaSame to accomplish a task, not just register an MCP?

The Agent Execution Router: a bounded public-capability router for AI agents, distinct from the MCP-owner registration flow covered in Getting Started.

Collection
documentation
Updated
2026-07-30
min read
6
Version
v1
Verified summary

Call resolve_and_run on the same public MCP endpoint with a goal, environment and constraints; it returns a bounded route and an honest execution state — only "verified" means something actually ran.

01

What this is for

Getting Started covers connecting an MCP to the Factory as an owner. This is the other direction: using SaSame's public MCP as an agent to get something done — search issues, find research, extract text, check whether a server is callable — via one router call instead of hand-picking a tool yourself.

02

The primary call

Call resolve_and_run on the same endpoint (https://live-vps.sasame.online/public-mcp, no auth) with:

{
  "goal": "Describe the outcome you need",
  "environment": "your client or runtime",
  "constraints": ["read only", "free only"],
  "execution_mode": "plan_only"
}

It resolves the goal into live candidate tools, ranks a primary route and fallbacks, and returns an honest execution state — never a fabricated success.

03

Execution modes

  • plan_only — returns the route and handoff without executing the downstream tool.
  • safe_read_only — executes only an explicitly requested, conservatively read-only public tool when the required arguments are present.

Neither mode writes, pays, or mutates anything.

04

Execution states — only "verified" is complete

StateCompleteMeaning
plannedNoA route exists, but nothing downstream executed.
handoff_requiredNoA route exists, but authorization, required arguments, or a human/owner action is still needed.
verifiedYesA permitted downstream call returned substantive output and passed the router's verification contract.
failedNoExecution was attempted but did not produce a verified result.
refusedNoThe request violated a safety, credential, payment, mutation, or input boundary.
no_routeNoNo sufficiently relevant observed public capability was available.

This matters because a router that reports discovery as delivery is dishonest by construction — planned and handoff_required are explicitly not success.

05

Hard boundaries

  • Public observations and grades are measurement status, not endorsement, safety, quality, malware, KYC, identity, sanctions, or legal-authority verdicts.
  • A plan, candidate route, or handoff is never reported as executed or verified.
  • The public router does not accept credential-like endpoint data or secret-like arguments.
  • AI agents execute autonomously by default inside honest, logged, verify-or-rollback envelopes — exposing owner shell/file tools publicly, spam, impersonation, anti-bot evasion, and committing secrets are prohibited.
  • SaSame is non-custodial and does not hold or move third-party funds.
  • tools/list and the live health contract are authoritative for the current callable surface — static copy never promises a fixed tool count.
06

Machine entrypoints

The same contract is available in multiple representations for different callers:

  • Router manifest: https://live-vps.sasame.online/.well-known/sasame-router.json
  • Agent card: https://live-vps.sasame.online/.well-known/agent-card.json
  • llms.txt: https://live-vps.sasame.online/llms.txt

Operated by SaSame SRL, Romania / European Union. The target audience is worldwide; jurisdiction is legal context only, not a geographic restriction.

EX

Examples

4
  1. 01

    Search GitHub issues related to an authentication bug — read only, free only.

  2. 02

    Find recent biomedical evidence about a treatment question — read only.

  3. 03

    Extract text from a public screenshot — read only, no credentials.

  4. 04

    Check whether a public MCP endpoint is callable before connecting it — measurement only.

RF

References

2