Model Context Protocol

Connect any AI client to GenZAgents.

One npm package. 21 MCP tools to issue receipts, run buyer lookups, capture conversations, handoff across providers, and proactively surface prior org work. All signed with your keys, all anchored to your verified identity.

Tested with

Claude Desktop
Claude Code
Cursor
Cline
Windsurf
Any MCP client

For Claude.ai and ChatGPT web (where MCP isn't available), use our browser extension or /import/openai bulk importer.

Pick the install path that fits.

Four ways to get the MCP server into every AI tool on your machine. All four land on the same config; pick whichever feels right.

Recommended · ~10 seconds↓

CLI installer

One npx command auto-detects every AI tool you have and merges the MCP config into each one. Always backs up the existing config.

Per tool · 2 minutes↓

Manual config snippet

See exactly what gets added before adding it. Copy-paste the JSON block into each tool's config file yourself.

Browser-friendly · click-through↓

Browser extension

A Chrome extension prompts you to install per detected tool. For people who don't run shell scripts. Native install — coming in v0.7.6.

Enterprise · IT-pushed→

MDM templates

JAMF, Intune, Workspace ONE config profiles. Push to fleets without each employee touching a terminal. Uses scoped install-only tokens.

Path 1

CLI installer — recommended

One npx command, ~10 seconds, configures every AI tool you have installed. Idempotent, backs up existing configs, requires no sudo, source-available on GitHub.

# 1. Set your API key in your terminal
export GENZAGENTS_API_KEY=sk_live_…

# 2. Run the installer (auto-detects Claude Desktop, Claude Code,
#    Cursor, Cline, Windsurf, Continue, Roo Code)
npx @genzagentsio/setup
What does this do, exactly?
  • Detects which AI tools are installed by checking canonical config paths
  • Reads each existing config, writes a .bak-<timestamp> backup
  • Merges mcpServers.genzagents into each config — preserves all other MCP servers + unrelated keys
  • Writes your API key to ~/.genzagents-mcp-env (mode 600). The literal key never appears in any AI tool config — each just references the env var
  • No sudo. No network calls except the npm install of itself. No telemetry.

Source: github.com/genzagents/genzagents/tree/main/packages/installer

--dry-run

Show what would change without writing.

--uninstall

Remove our MCP entry from every detected tool.

--silent --yes

For MDM / CI — no prompts, single-line summary.

Path 2

Manual config snippet — see exactly what gets added

Three steps. Less than two minutes per AI client.

1

Get an API key

Sign in at genzagents.com → Settings → API keys → Generate key. Copy it once — you only see the secret on creation.

sk_live_...
2

Drop the config into your AI client

Same snippet works across every MCP-compatible client. File locations:

  • Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows)
  • Claude Code: ~/.claude.json or per-project .mcp.json
  • Cursor: ~/.cursor/mcp.json
  • Cline / Windsurf: Settings → MCP Servers → Add custom server
{
  "mcpServers": {
    "genzagents": {
      "command": "npx",
      "args": ["-y", "@genzagentsio/mcp-server"],
      "env": {
        "GENZAGENTS_API_KEY":     "sk_live_...",
        "GENZAGENTS_AGENT_DID":   "did:genz:...",
        "GENZAGENTS_AUTO_RECEIPT": "tool",
        "GENZAGENTS_MCP_PROJECT":  "My agent's main chat"
      }
    }
  }
}

Tip: pin a version (e.g. @genzagentsio/mcp-server@0.6.0) for deterministic production builds.

3

Restart your client and verify

Quit and relaunch so it picks up the new config. Then ask your agent:

"What MCP tools do you have for genzagents?"

You should see all 21 tools listed. If not, check the client's MCP logs (Claude Desktop: Settings → Developer → MCP).

Path 3

Browser extension — click-through install

For people who don't want to run shell scripts. The extension already captures Claude.ai web + ChatGPT web conversations as receipts; in v0.7.6 it also detects locally-installed Claude Desktop / Cursor / Cline via native messaging and offers a click-through install prompt per tool.

Today: The extension captures Claude.ai web + ChatGPT web conversations as receipts on your account. The auto-install-into-Claude-Desktop part lands in v0.7.6 (~1 week post-launch).

Source on GitHub →

Want it sooner? Email hello@genzagents.io — we'll prioritise based on demand.

Path 4

MDM templates — push to a fleet

JAMF, Intune, Workspace ONE config templates that push the installer to every employee's laptop using a scoped install-only token. The token can configure tools on disk but cannot read receipts, issue receipts, or authenticate to anything else. After install each employee signs in via SSO and gets their own per-user identity.

See MDM templates →

21 tools, six categories

Every call is authenticated with your API key and scoped to what your org can see. No data leaves your control — the MCP server is a thin remote-call wrapper, not a state store.

Identity4 tools
  • list_my_agentsList every agent registered to you — DIDs, capabilities, trust scores, receipt counts.
  • get_agentLook up any agent by DID. Returns name, bio, capabilities, owner, environment, trust score.
  • get_ownerFetch the verified human behind an agent.
  • get_my_trust_scorePull your own trust score breakdown.
Receipts6 tools
  • draft_receiptSign and submit a draft receipt as a buyer.
  • countersign_receiptSign as the seller agent to finalise a receipt.
  • verify_receiptOffline verification of both signatures and canonical hash.
  • list_my_receiptsPaginate receipts for one of your agents — filter by project, provider, env, date.
  • dispute_receiptOpen a dispute on a finalised receipt.
  • check_duplicatesPre-issue probe — before drafting, check if similar work already exists in your org.
Buyer3 tools
  • lookup_agent_reputationTrust score, KYC tier, recent categories, dispute rate for any DID.
  • watchlist_addTrack an agent and get alerts on score moves or new disputes.
  • generate_evidence_packBuild a SOC 2 / ISO 42001 / EU AI Act / EU CRA compliance pack.
Lifecycle4 tools
  • snapshot_agent_memoryCapture an agent's memory state — sources: Mem0 (live), Letta, Anthropic transcript, manual paste.
  • list_agent_snapshotsList every memory snapshot for one of your agents, newest first.
  • request_agent_accessRequest scoped access to another owner's agent.
  • transfer_agentCryptographic transfer of ownership to a new verified human (with key rotation).
Continuity3 tools
  • list_my_chatsHierarchical view: provider → project → recent receipts. Used by restore_chat to find a target.
  • restore_chatOne-click cross-provider handoff. Returns system prompt + receipt digest + memory snapshot URL for the host LLM to ingest.
  • org_context_lookupPROACTIVE org-wide de-dup. Call FIRST at conversation start with the user's first message; returns relevant prior org work so you don't redo what a colleague already did.
Capture1 tool
  • capture_sessionSelf-issue a receipt for the conversation content itself (auto-receipt covers tool-call actions; this captures the discussion).

What the MCP server does — and doesn't — see

  • ✓Runs locally as a Node process. Communicates with api.genzagents.com over HTTPS only.
  • ✓Auto-receipt mode tags every receipt with the conversation's project + provider + per-session UUID — so the receipt naturally buckets per chat.
  • ✓API key lives in your AI client's config (or the side-file at ~/.genzagents-mcp-env). Revoke anytime at /settings/api-keys.
  • ✗Doesn't see your filesystem, your other apps, or anything outside the GenZAgents API.
  • ✗Doesn't hold private keys. Self-issued receipts are signed server-side with our issuer key; party-signed receipts use your account's registered keypair.

Troubleshooting

Tools don't show up in my client+

Fully quit and relaunch the client (not just close the window). Check the client's MCP logs (Claude Desktop: Settings → Developer → MCP) for connection errors. Most common cause: a typo in the config JSON — most clients silently ignore invalid MCP entries.

Calls return 401 Unauthorized+

Your GENZAGENTS_API_KEY is wrong, expired, or revoked. Generate a fresh one at /settings/api-keys and update your config.

Calls hang or time out+

Network or DNS issue. Try `curl -s https://api.genzagents.com/health/ready` from a terminal — you should get JSON with postgres:true.

Where do I see what my AI did?+

Every tool call that mutates state shows up at /dashboard — receipts on the Receipts tab, agent edits on My Agents, watchlist changes under Verify Agents. The MCP server itself doesn't keep history; the API does.

How do I enable auto-receipts?+

Set GENZAGENTS_AUTO_RECEIPT="tool" in the MCP env block. Every non-readonly tool call now auto-issues a self-signed receipt tagged with the conversation's project + provider + per-session UUID. Use "session" mode for one summary receipt at chat end instead of per-call.

Can I run this without exposing the API key in claude_desktop_config.json?+

Yes — drop the env vars into ~/.genzagents-mcp-env (mode 600) instead. The MCP server boot loads this side-file even when Claude Desktop overwrites its own config on save.

Ready to wire up your agent?

Get an API key, drop the snippet into your client, restart. Two minutes. Free MCP, free tier covers 1k receipts/month.

Get API key →Source on GitHub