What is Qoc?
An overview of Qoc — the self-hosted AI trading desk driven by the coding agents you already use.
Qoc is a self-hosted trading desk that lets AI coding agents — Claude Code and any MCP-compatible agent — research markets, propose trades, and execute orders on your behalf, with you as the final approver on every decision.
A desk you own and control
Qoc runs entirely on your own machine or server. There is no cloud intermediary holding your positions, reading your strategy, or routing your orders. Your workspace is a directory of plain, human-readable files tracked under built-in version history — you can inspect, edit, or roll back any file at any time.
Everything the desk knows — your holdings, research notes, risk rules, schedules, and order history — lives in that directory. You can commit it to your own private git repository, back it up however you like, and carry it between machines without a migration wizard.
Who Qoc is for
Qoc is built for traders and engineers who are already comfortable working with AI coding agents and want to apply that same workflow to financial markets. If you can write a GitHub issue and review a pull request, you already understand how Qoc works.
It fits active individual traders who want research automation, systematic traders who want agent-driven strategy execution, and developers who want a scriptable, programmable desk they can extend without a vendor SDK.
The research → decide → execute loop
Every action the desk takes follows a three-stage loop. In the research stage the agent pulls quotes, fundamentals, news, and technicals into your workspace, updates entity files for every ticker you track, and writes or revises a thesis in research/.
In the decide stage the agent evaluates the research against your risk rules (desk.toml guards) and, if a trade is warranted, writes a proposed order file into orders/proposed/. That file is a plain TOML document you can read, edit, or delete.
In the execute stage you review and approve the proposal — or reject it with a comment. Once approved, the desk routes the order through the connected venue and records the fill in inbox/ and orders/executed/. Nothing executes without your explicit approval.
Every trade is a proposal
Qoc never sends an order without a human-approved proposal file. Think of each proposal as a commit you review before it is merged. The agent writes it; you merge it. Guards enforce hard limits before the order ever reaches a venue.
Why a coding agent runs your desk
Coding agents are already expert at reading documents, writing structured output, calling APIs, and working inside a version-controlled directory. Those are exactly the skills a trading desk needs. Qoc gives the agent a well-defined workspace layout, a set of CLI commands (qoc run, qoc status, qoc snapshot), and an MCP server that exposes market data and order management as tool calls.
You interact with the desk the same way you interact with any agent task: open an issue, describe what you want to research or trade, and let the agent work through it. The agent's reasoning is written into files you can read and audit — not locked inside a chat session.
A typical agent session
# Start the local desk service
qoc up
# Hand the agent a task described in issues/001-earnings-review.md
qoc run issues/001-earnings-review.md
# Check what the agent proposed
qoc status
# Review the proposal and approve it
cat orders/proposed/20260705-001.toml
qoc approve orders/proposed/20260705-001.toml