Skip to main content
Glama
Teller Connector Demo

Teller Connector Demo

by Aly700

Teller connector demo

An MCP connector that lets an agent read accounts and request transfers in Teller, a simulated payments API.

Demo data. No real money.

The agent gets three read tools and one write tool. It can look up accounts and transfers and ask for a transfer, but it cannot approve one. Every transfer request waits on a web page until a person approves it there, and Teller gets nothing before that. Approve, and the transfer posts once. Reject, and nothing is sent. If something crashes mid-transfer, it still never posts twice: a request with an unknown result is marked uncertain until it is checked.

What is in the repo

  • MCP tools: list_accounts, get_account, get_transfer and request_transfer.

  • Approval console with a live request inbox, account balances, event timelines, deliberate browser approval, rejection and recovery. Inspect the session audit or download JSON lines. Dark and light themes follow your system preference, with a manual override.

  • 80 evaluation cases and a GitHub Actions workflow that uses your own Anthropic key. The report says "Not run yet" until model results are recorded.

  • Hosting config for Cloudflare Workers, with Teller and Postgres on a small, always-on Mac behind a Cloudflare Tunnel.

  • Teller source in vendor/teller, plus demo data, database migrations and tests.

Related MCP server: Parimit

Run it locally

You need macOS, Node 26 and JDK 25 on PATH, and Homebrew PostgreSQL 16. From the repo root:

npm ci
scripts/start-local.sh

Open http://localhost:8787 (use localhost, not an IP address). Connect an MCP client to http://localhost:8787/mcp and choose Start demo session. Each start resets the demo data. Local state, secrets and logs stay in .data/. You need no model key or cloud account to run it.

To stop everything:

scripts/stop.sh

To run the evaluation, create a GitHub environment called model-eval with the secrets ANTHROPIC_API_KEY and EVAL_REPORT_SECRET and the variables EVAL_MODEL_ID and EVAL_REPORT_ORIGIN (the address of your report page). It stays off until you set the repository variable ENABLE_MODEL_EVAL=true. To run locally in a dedicated evaluation checkout, use EVAL_DATASET=isolated npm run eval:ci; it replaces that checkout’s demo database.

Tests

Install Google Chrome, then run:

scripts/ci.sh

It installs everything, builds Teller, and tests approval, rejection, crash recovery and reset in a real browser. It runs on this machine only and calls no model. Set NODE_BIN, JAVA_HOME, PG_BIN or CHROME_PATH to use other installs. Teller's own Docker-based tests are not run.

The console brings requests, balances, policies, audit events, evaluation cases and connection settings into one workspace, with measured session charts and plain-language request explanations. Server-rendered pages use one dependency-free script for search, tables, request drawers, keyboard controls, time display and Auto, Light or Dark themes. Approval also works without JavaScript.

License

MIT. See LICENSE. Teller is also by Affan Yasir and is included unmodified from commit 9418e33, under the same licence. The bundled fonts, Manrope and IBM Plex Mono, are under the SIL Open Font License; notices are in apps/mcp/assets/FONT-LICENSES.txt.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to propose structured payment intents, apply deterministic policy, require human approval, and record tamper-evident audit trails without connecting to live payment rails.
      MIT
    • A
      license
      Not graded
      quality
      C
      maintenance
      Enables natural-language-driven purchase-to-pay workflows against mock procurement and payment APIs, including invoice matching, exception handling, human approvals, and idempotent payment scheduling with a tamper-evident audit log.
      MIT
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to request policy-governed payments through MCP tools, with server-side enforcement of spending limits, merchant allow-lists, Pix BR Code validation, approval thresholds, and hash-chained audit logging in a sandbox. It lets agents check wallets, create card or Pix payment intents, and poll or list payments while keeping approval and limit changes out of agent control.
      Apache 2.0