# TRKR agents

Connect your own AI assistant to TRKR's synthetic market data.

TRKR simulates hundreds of complete alternative futures for real assets — every
one a full ten years of daily prices, anchored to today's market and generated by
a regime-aware, macro-driven engine calibrated against two decades of history.
This plugin lets Claude query that data directly; ChatGPT and any other MCP
client can reach the same endpoint (see below).

**The simulations stay on TRKR's servers; the thinking happens in the assistant
you already pay for.** TRKR sends small, reduced answers — quantile bands, summary
statistics, downsampled series — never the raw grid. That keeps your context
small and your bill low.

## Install (Claude)

```
/plugin marketplace add trkr-ai/trkr-agents
/plugin install trkr
```

Then set your API key, which you can create at
[trkr.ai/app/agents](https://www.trkr.ai/app/agents):

```bash
export TRKR_API_KEY=trkr_sk_...
```

Or, without the plugin, point any MCP client at the endpoint directly:

```bash
claude mcp add --transport http trkr https://www.trkr.ai/api/mcp \
  --header "Authorization: Bearer $TRKR_API_KEY"
```

## ChatGPT

ChatGPT cannot send an API key to an MCP server — OpenAI's connectors support
OAuth only — so it signs you in instead. There is no key to copy.

1. Turn on [developer mode](https://help.openai.com/en/articles/12584461-developer-mode-and-mcp-apps-in-chatgpt)
   in ChatGPT's settings. OpenAI has moved this between **Security and login**,
   **Connectors → Advanced settings** and **Apps → Advanced**, so follow their
   help article rather than a path written here. It is on chatgpt.com only (not
   the mobile app), needs a paid plan, and on a Business or Enterprise workspace
   an admin must allow custom connectors first.
2. Create a connector pointing at `https://www.trkr.ai/api/mcp`, authentication
   **OAuth**.
3. You will land on trkr.ai to sign in and approve the connection, then return
   to ChatGPT with the tools available.

The skill in this repo is a Claude format, so a ChatGPT connector does not
install it. The same rules travel in the server's own `instructions`, which
ChatGPT reads at connection time — but if you use TRKR in a ChatGPT Project,
pasting `trkr/skills/reading-synthetic-markets/SKILL.md` into the project's
custom instructions is the closest equivalent.

## Google Antigravity

Gemini CLI was retired in June 2026 and replaced by **Antigravity** (the `agy`
command). The consumer app at gemini.google.com still has no way to add a
connector, so Antigravity is the route for Google users — and it is free.

```bash
curl -fsSL https://antigravity.google/cli/install.sh | bash
agy install
```

Then, one command — flags must come before the name, and an http URL is
detected without `--type`:

```bash
agy mcp add --header "Authorization: Bearer $TRKR_API_KEY" \
  trkr https://www.trkr.ai/api/mcp
```

Or write it to `~/.gemini/config/mcp_config.json` yourself (`.agents/mcp_config.json`
for one workspace). The field is `serverUrl`; Antigravity rejects the `url` and
`httpUrl` spellings Gemini CLI used, and the wrong one fails as though the server
were down:

```json
{
  "mcpServers": {
    "trkr": {
      "serverUrl": "https://www.trkr.ai/api/mcp",
      "headers": { "Authorization": "Bearer trkr_sk_..." }
    }
  }
}
```

Drop the header entirely and Antigravity handles OAuth by itself — it registers
dynamically and sends you to trkr.ai to approve, the same flow ChatGPT uses.


## What you can ask

- *"Why does future #42 crash in 2031?"* — it reads the regimes, macro paths and
  prices behind that specific simulation.
- *"Critique 60/30/10 SPY/AGG/GLD."* — run through every future at once, with the
  outcome distribution rather than a single backtest.
- *"Build something for a client retiring in five years."* — co-build the
  allocation, then test it.
- *"Is this data credible?"* — the quality scorecard, per asset, rather than an
  assurance.

## What's in here

| | |
|---|---|
| `trkr/.mcp.json` | The MCP server: seven tools over the current run |
| `trkr/skills/reading-synthetic-markets/` | How to read an ensemble without drawing false conclusions — the spread vs the median, unit conventions, and why filtering futures by realized returns invalidates a benchmark comparison |

The skill matters as much as the tools. Simulated ensembles invite a specific
class of confident, wrong conclusion, and the skill is what stops your assistant
producing them.

## Not investment advice

TRKR is simulated data for analysis and stress testing. No output is a prediction
about real markets. Methodology: [trkr.ai/methodology](https://www.trkr.ai/methodology).