AI agent (any host).

The classic — an MCP-powered agent with the full 32-tool Sera surface. Three paths depending on how custom you want to go: bundled Sera Agent (one command), Claude Code/Cursor/ChatGPT (just register the MCP), or your own host using OpenAI Agents SDK.

Path A — Bundled Sera Agent (fastest)

Best if you just want a working agent now and don't care which host it runs in. Comes pre-wired with MCP + LLM.

git clone https://github.com/Josh-sera/sera-agents
cd sera-agents/sera-agent && npm install
export OPENAI_API_KEY=sk-...
npm start

# Try:
> What stablecoins do you support for SGD?
> How much USDC to deliver exactly 5,000 MYR?
> Run sera.doctor

Path B — Claude Code (or Cursor, ChatGPT, etc.)

Best if you already use a modern MCP host. Three lines of config, the host loads the MCP at startup, the agent gets all 32 tools natively.

1Get sera-mcp on your machine

git clone https://github.com/Josh-sera/sera-mcp
cd sera-mcp && npm install && npm run build

2Register with the host

For Claude Code:

claude mcp add sera --scope user \
  --env SERA_NETWORK=mainnet \
  --env POLICY_PRESET=standard \
  -- node $(pwd)/dist/index.js

For Cursor: Settings → MCP → Add a server with the same command. For ChatGPT: Settings → Connectors → MCP. See integrations/ for per-host snippets.

3Verify

In any agent session, type:

Call sera.doctor

You should get back overall_ok: true with green checks for sera_health, network_sanity, tokens_registry, signer_mode, policy.

Path C — Your own host (OpenAI Agents SDK)

Best if you're building a custom agent product and need programmatic control over the loop.

import { Agent, MCPServerStdio } from '@openai/agents';

const sera = new MCPServerStdio({
  command: 'node',
  args: ['/abs/path/to/sera-mcp/dist/index.js'],
  env: { SERA_NETWORK: 'mainnet', POLICY_PRESET: 'standard' }
});

const agent = new Agent({
  name: 'fx-helper',
  instructions: 'You help with multi-currency FX settlement. Use sera.doctor first if anything looks off. Always show cost in basis points and absolute dollars.',
  mcpServers: [sera]
});

const response = await agent.run('Find me FX deals over 25 bps right now.');
console.log(response);

Add policy gates (any path)

For any agent that will execute swaps, set the policy preset and caps via env vars. sera-mcp reads these and refuses to exceed them — even if the LLM tries.

SERA_NETWORK=mainnet
POLICY_PRESET=standard           # or 'strict' or 'dry-run-only'
SERA_DAILY_VOLUME_CAP_USD=10000  # server-derived, not LLM-claimed
SERA_DRY_RUN=false               # true = every execute returns dry-run trace
SERA_HISTORY_DB=$HOME/.sera/history.db

Useful slash commands

The MCP exposes 4 pre-templated prompts that hosts surface as commands:

• /deal_scan — find FX deals across all pairs above an edge threshold
• /treasury_brief — multi-currency exposure report + drift
• /invoice_optimizer — cheapest source for a given target payout
• /fx_integrity_check — triangular-arb scan across a fiat basket

Common gotchas

• The agent claims it executed when it didn't. Hosts vary in how they show tool-call results. Always trust sera.doctor + sera.settlement_status over the LLM's narrative.
• Quote expiration. Quotes are short-lived (~30s). If the agent dawdles between get_quote and execute_swap, fetch a fresh quote.
• Default to simulate:true when exploring. Pass it on get_quote to probe without a wallet — saves friction during development.
• Long contexts confuse tool-routing. If the agent stops calling sera.* tools mid-conversation, it might have lost the system prompt. Re-anchor with a fresh "use the sera tools to..."

Done with the tutorials. Next: browse the full API reference, copy a recipe, or read about how the layers compose.