Tally — anomaly detection for AI trading agents

Every trading connector makes your AI more capable. Tally makes it accountable.

Tally (tally.markets) is anomaly detection for the AI agent that trades a brokerage account. It reads the account read-only through SnapTrade, builds the agent's own baseline (symbols, size, frequency, hours, cancels), scans the account against it every hour, and emails the owner the hour the agent deviates. Prompt injection, a poisoned page, a bad update: a compromised agent shows up first as an order it would never have placed yesterday, and that is what Tally flags. Most days it sends nothing.

This connector is how an AI running inside Claude, ChatGPT, Claude Code, or any MCP-capable client takes part in that. Four things it does:

• The AI reads the monitoring (get_checkup): the status word per account, orders in the last 24 hours, anomaly alerts of the last 7 days and whether the owner answered them, the broker connection. The morning checkup leads with it.
• The agent declares its mandate (declare_mandate): what it trades, the largest order as a share of the account, cadence, hours, usual symbols, as structure. Tally runs it from the first scan and never loosens it; a later change notifies the owner.
• The agent reports what it researched, considered, and proposed (log_agent_activity): symbols and kinds only, no text accepted or stored. Tally judges the account by its actual orders; the reports sharpen an alert (an order in a symbol the agent never mentioned) and never replace one.
• The record, built from the broker's orders without anyone typing: every completed round trip with its realised P&L, and the numbers (win rate, expectancy, profit factor, by symbol, setup, weekday). The agent's setup tag on a proposal lands on the trip.
• Pre-registered theses, optional, for trades the user makes themselves: a statement, exit rules that freeze at creation, an optional 24-hour cooling-off, an override ledger for rule-breaking exits, and a public receipt for every closed trade.

Tally never executes trades. Every broker and venue connection is read-only at the API level — no order placement, no custody, by architecture and by policy. Nothing in this connector can reach a brokerage.

Connect

Server URL (Streamable HTTP · OAuth 2.0 with dynamic client registration):

https://tally.markets/api/mcp

Client	How
Claude	Settings → Connectors → Add custom connector → name it Tally, paste the URL. Leave the OAuth fields empty — Tally registers itself.
ChatGPT	Listed app — one click from the directory, no developer mode. (Manual fallback: developer mode → Plugins → New plugin, the URL as Server URL, OAuth.)
Claude Code	claude mcp add --transport http tally https://tally.markets/api/mcp, then /mcp to sign in.
Any MCP agent	Point it at the URL above; OAuth discovery does the rest.

Then say: "Get started with Tally, then run my checkup." The server teaches your AI the whole protocol on first contact. The connector needs a subscribed Tally account (Sentinel, $4.99 a month or $49 a year); every tool but get_started answers only for one.

Full walkthrough (skill, scheduled morning checkup, broker linking): tally.markets/connect

Tools

Tool	What it does
get_started	Returns the protocol — the AI teaches itself
declare_mandate	Declare what the agent is supposed to do — asset classes, largest order, cadence, hours, symbols; a change notifies the owner
log_agent_activity	Report what the agent researched, considered, or proposed — symbols, kinds, and a setup tag; no text
get_checkup	Morning briefing: the monitoring status per account and any anomaly alert awaiting an answer, then active theses, deadlines, evidence, closed history and lifetime R
list_theses / get_thesis	Read the journal
create_thesis	Pre-register a trade: statement, mechanism, risk budget, exit criteria; armed at once, or cooling_off: true for the 24h gate
arm_thesis	Arm a thesis whose cooling-off has passed — from here the criteria fire, never edit
record_reading	Log evidence against a criterion (the generic sensor primitive)
fire_criterion	Mark a rule objectively triggered, with cited evidence — one-way
close_thesis	Close a trade; discretionary closes require a written justification, logged forever
get_receipt_link	The public receipt for a closed thesis

Reads are safe to always-allow; writes are guarded by the protocol itself (cooling-off, criteria minimums, override justifications).

Prompts

Clients with a prompt picker get the three jobs Tally exists for, ready to run:

Prompt	What it does
morning-checkup	The daily run — the monitoring status and any anomaly alert awaiting an answer first, then fresh readings against every active thesis's criteria and anything that fired. This is the one to put on a schedule.
pre-register-trade	Turns an idea into a pre-registered thesis, interviewing you until it is falsifiable
close-out	Walks a thesis to its exit — what the written rules demand, and the justification the override ledger requires if you are closing early

Proof

A real receipt — pre-registered thesis, venue-verified entry and exit, −0.22R, graded Process B / Outcome D, closing note verbatim: "Wrong on thesis, right on process."

How it compares

Broker notifications confirm each fill. The agent's own reports grade the agent. Portfolio trackers show the balance. Tally is the one built to check the account against the agent's own baseline, independently of all three: tally.markets/compare

Privacy & security

• OAuth per user; your AI sees your own record only, under a grant you can revoke any time.
• What the agent reports is stored as a symbol, a kind, and a time, for 90 days. No text field exists.
• Brokerage data (via SnapTrade) is read-only and is read by Tally's own service, never through this connector; broker credentials never touch Tally.
• Details: privacy · terms · support

---

Tally is a hosted service. This repository is its documentation and public manifest, not its source — there is nothing here to install or run. Point your AI at the connector URL above; OAuth does the rest.