Skip to main content

Claude Managed Agents

Add geospatial identity, jurisdictional binding, and EU AI Act compliance to Claude Managed Agents in one SSE endpoint — no infrastructure required.

Why GEIANT + Claude Managed Agents​

Claude Managed Agents gives you the execution runtime. GEIANT gives you the accountability layer. Together they form the first complete sovereign agent deployment stack:

LayerProviderWhat it does
ExecutionAnthropicSandboxed runtime, long-running sessions, multi-agent coordination
IdentityGEIANTEd25519 keypair — the agent's cryptographic identity
JurisdictionGEIANTH3 territorial binding — which laws apply
Audit trailGEIANTSigned breadcrumb chain — what the agent did, when, under whose authority
ComplianceGEIANTAuto-generated EU AI Act Art. 12 / Art. 14 report

Connect​

Add GEIANT to your Claude Managed Agents session via the mcp_servers parameter:

const response = await fetch("https://api.anthropic.com/v1/messages", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    "anthropic-beta": "managed-agents-2026-04-01",
  },
  body: JSON.stringify({
    model: "claude-sonnet-4-20250514",
    max_tokens: 1000,
    mcp_servers: [
      {
        type: "url",
        url: "https://packagesmcp-perception-production.up.railway.app/sse",
        name: "geiant",
      }
    ],
    messages: [
      { role: "user", content: "Analyse flood risk for Rome and generate a compliance report." }
    ],
  }),
});

Every tool call made through the GEIANT MCP server automatically:

  1. Pre-flight — checks the delegation certificate (jurisdiction, facet, temporal validity)
  2. Executes — runs the tool
  3. Post-flight — builds a SHA-256 context digest, chains to the previous block, Ed25519-signs, writes to the audit trail

If pre-flight fails, the tool call is blocked — it never executes.

Available Tools​

Perception​

ToolDescription
perception_weatherHistorical / forecast weather for any H3 cell via ERA5
perception_fetch_tileSentinel-2 L2A satellite tiles via Element84
perception_classifyPrithvi-EO-2.0 foundation model inference (flood detection, land use)
perception_embedClay v1.5 geospatial embeddings (coming soon)

Governance​

ToolDescription
gns_get_compliance_reportFull EU AI Act Art. 12 / Art. 14 report with chain verification, epochs, delegation cert, and violation history
gns_get_trust_scoreCurrent TierGate tier and trust score
gns_verify_chainCryptographic integrity check of the breadcrumb chain
gns_roll_epochSeal pending breadcrumbs into a Merkle-rooted epoch snapshot

Example Agent Session​

A Claude Managed Agent analysing flood risk and producing a compliance report in a single session:

// The agent autonomously calls tools in sequence:

// 1. Fetch satellite tile for Rome
perception_fetch_tile({ h3_cell: "851e8053fffffff", bands: ["B04", "B8A"] })

// 2. Run flood detection
perception_classify({ h3_cell: "851e8053fffffff", task: "flood_detection" })

// 3. Get current weather context
perception_weather({ h3_cell: "851e8053fffffff" })

// 4. Generate compliance report mid-session
gns_get_compliance_report()
// → {
//     agent_pk: "c14094ea...",
//     agent_handle: "energy@italy-geiant",
//     current_tier: "provisioned",
//     trust_score: 20.99,
//     chain_verification: { is_valid: true, block_count: 8, issues: [] },
//     violations: []
//   }

// 5. Seal the session into an epoch
gns_roll_epoch()
// → {
//     epoch_index: 1,
//     merkle_root: "6fa3e35a...",
//     block_count: 3
//   }

Every one of those tool calls is now a signed, hash-chained breadcrumb in the audit trail — cryptographically traceable to the human principal who authorized the agent.

EU AI Act Compliance​

GEIANT generates runtime evidence for the two articles that documentation-only compliance cannot satisfy for agentic AI:

Article 12 — Record-keeping Every tool call automatically produces a signed breadcrumb with:

  • context_digest — SHA-256 of tool input + output
  • identity_public_key — Ed25519 agent identity
  • delegation_cert_hash — link to the human authorization
  • Hash chain — tamper-evident, verifiable offline

Article 14 — Human Oversight The delegation certificate proves:

  • Which human principal authorized the agent (principal_pk)
  • What the agent is allowed to do (facets, allowed_tools)
  • Where it is allowed to operate (h3_cells)
  • For how long (not_before / not_after)
  • How many levels of sub-delegation are permitted (max_depth)

Pull the full report at any time:

curl https://packagesmcp-perception-production.up.railway.app/compliance | jq .

Structural Advantage over Existing Solutions​

Microsoft Entra Agent ID cannot produce:

  1. Third-party verifiable proof of jurisdiction — Ed25519 signatures are verifiable by anyone without contacting a central authority
  2. Geospatially-bound delegation — H3 cells are a geographic primitive, not an organizational one
  3. Cryptographically immutable operation log — breadcrumb chains cannot be silently modified; any tampering invalidates all subsequent blocks

This is GEIANT's structural moat for EU AI Act compliance — and it plugs directly into Claude Managed Agents with a single SSE URL.

Live Status​

curl https://packagesmcp-perception-production.up.railway.app/health | jq .
{
  "status": "ok",
  "service": "geiant-mcp-perception",
  "version": "0.3.1",
  "audit_active": true,
  "agent_pk": "c14094ea7efb6122...",
  "chain_tip": 8,
  "tools": [
    "perception_fetch_tile",
    "perception_classify",
    "perception_embed",
    "perception_weather",
    "gns_get_compliance_report",
    "gns_get_trust_score",
    "gns_verify_chain",
    "gns_roll_epoch"
  ]
}

See Also​