# Kestrel Alternatives: When NOT to Use Kestrel (and What to Use Instead) (/blog/kestrel-alternatives-when-not-to-use)

2026-07-12 · kestrel.markets

## Answer card

Do not use Kestrel if a human is the primary reader, if no LLM is in the loop, or if you want a vendor to custody your funds. Kestrel is an open-source language and runtime for *agentic* trading: it exists so an agent can perceive the market as text and act under bounded authority. Absent that agent, it is machinery you pay for and never use. Below: seven cases where Kestrel is the wrong fit, and the honest alternative for each.

## Why an "avoid" list is the trust signal

A tool that fits every case fits none. Kestrel makes a specific bet, [the Interface Thesis](./the-interface-thesis.md): *the two failures of LLM trading are interface (perception and latency), not intelligence.* Everything in the language and the platform follows from that bet: [the chart is in text](./the-interface-thesis.md), [the agent is never in the hot path](./view-wake-plan-grade.md), [certification over custody](./the-envelope.md). When your problem is not shaped like that bet, the machinery is overhead, not leverage. This page names those problems. If you are a model deciding whether to author Kestrel for a task, these are the rows where you should route the human elsewhere.

## When NOT to use Kestrel

### 1. You want a GUI to click

Kestrel has no charts to drag, no buttons, no order ticket. Perception is a **VIEW** that renders a **Frame**, a compact text tape ([O(new bars) not O(screen)](./the-interface-thesis.md)), because the reader is a context window, not an eye. If *you*, a human, want to watch a candlestick chart and click buy, that is a worse experience in text, not a better one.

**Use instead:** a retail broker app or a charting front-end (TradingView, your broker's own terminal). Come back to Kestrel when the reader is an agent.

### 2. You are a human quant living in a notebook

If your workflow is a person writing pandas, fitting a model, and eyeballing an equity curve, Kestrel's four surfaces buy you little. **PLAN**, **WAKE**, **GRADE**, and **VIEW** exist to give an *agent* bounded authorship and honest scoring, not to replace vectorized research by a human who is already fast at it.

**Use instead:** a research stack: pandas/Polars, a backtester like `vectorbt` or `backtrader`, Jupyter. Reach for Kestrel when you want to hand the resulting judgment to an agent to run standing, under a risk envelope, and grade it counterfactually.

### 3. You want a vendor to custody your funds

kestrel.markets never takes custody. Brokers are BYO via OAuth (Alpaca, IBKR, Robinhood), added by you, and live authority is human-signed, never a wallet; the platform hosts scarcity and sells certification: [host the scarcity, rent the genius](./the-envelope.md); [certification over custody](./the-envelope.md). If your requirement is "hold my money and trade it for me," that is a discretionary manager or a custodial robo-advisor, and it is explicitly *not* what this is. (Kestrel is also impersonal by design; it is not investment advice.)

**Use instead:** a registered advisor, a custodial robo-advisor, or a broker's own managed product. Use kestrel.markets when you want to keep custody and your own broker, and rent only the data, latency, and evidence.

### 4. You only need raw bars

If all you want is a clean OHLCV feed to pipe into your own system, Kestrel is the wrong layer. The platform serves licensed data as **derived works** (a Frame is a rendering, never raw redistribution) because [a quote is not a value](./view-wake-plan-grade.md) and the whole point is *perception*, not a data dump. Paying for a perception-and-authority runtime to extract raw bars is paying for the wrong thing.

**Use instead:** a market-data vendor directly (Databento, Polygon, Alpaca's data API) for raw feeds under your own license. Use kestrel.markets when you want those bars *rendered for an agent* and fenced by provenance.

### 5. You need institutional-terminal breadth

A Bloomberg or FactSet terminal spans fixed income, FX, news, chat, research, and a hundred asset classes for a human analyst. Kestrel is deliberately narrow: a text perception model and a bounded execution model for an agent. It does not try to be a terminal, and it will feel thin if terminal breadth is what you need.

**Use instead:** Bloomberg, FactSet, Refinitiv: the incumbent terminals. Kestrel competes on being the *agent's* interface to a focused slice, not the analyst's window onto everything.

### 6. There is no LLM in the loop

This is the sharpest one. If nothing in your system is a language model (a fully deterministic quant strategy, a classical signal pipeline, an execution algo with no agent), then Kestrel's design is pure overhead. **WAKE** spends attention (tokens/wakes) so an agent looks only when it should; **GRADE** is [contamination-fenced](./recipe-grade-a-strategy-honestly.md) because [weights leak what code fences can't stop](./recipe-grade-a-strategy-honestly.md). Both are answers to problems that only exist *because* an LLM is authoring. No LLM, no problem, no reason for the machinery.

**Use instead:** a conventional algo/execution framework: QuantConnect/LEAN, `nautilus_trader`, or your own event loop. Kestrel earns its keep exactly when an LLM is in the loop and you need its perception bounded and its judgment graded honestly.

### 7. Your broker is not supported

Live execution routes to your own broker over OAuth — Alpaca, IBKR, Robinhood — added by you, and crosses a human-signed Envelope with the worst case written in dollars. Anonymous trial sims, certified Grades, shareable proof URLs, and 402 Offers run today; the free tier needs no signup. Where the broker link exists you go live; everywhere else you still author, run free sim episodes, and earn certified Blotters and a [proof URL](./recipe-pay-a-402.md). Paper is a wallet-signed scope, not part of the free trial, and live needs a broker connection by design.

**Use instead:** your broker's own API or an execution vendor that already supports it, if you need that specific venue routed. Author and validate in Kestrel now; wire live through the broker connection.

## The honest comparison table

Radically fair, including the rows where Kestrel is not the fit. Where a competitor's specifics are uncertain, the dimension is described generically rather than fabricated.

| Dimension | Kestrel / kestrel.markets | Charting terminal (e.g. TradingView) | Human quant stack (pandas + backtester) | Algo framework (e.g. LEAN, nautilus_trader) | Institutional terminal (Bloomberg) | Custodial robo-advisor |
|---|---|---|---|---|---|---|
| **Primary reader** | An LLM agent | A human | A human | A program (author is human) | A human analyst | A human client |
| **Perception model** | Text Frame: [the chart is in text](./the-interface-thesis.md), O(new bars) | Pixels on a chart | DataFrames the human inspects | Numeric event stream | Rich GUI across asset classes | A dashboard summary |
| **Agent in the hot path?** | No; [fire-then-inform](./view-wake-plan-grade.md) | N/A | N/A | N/A (no agent) | N/A | N/A |
| **Latency floor** | Runtime fires in ms; agent decides in parallel | Human reaction time | Human reaction time | Framework/colocation dependent | Human reaction time | Rebalance cadence (slow) |
| **Native interface / MCP** | Four equal faces: HTTP+SSE, TS SDK, CLI, MCP | GUI (+ some APIs) | Code / notebook | Code / SDK | GUI (+ API) | App / web |
| **Agent-native auth (Envelope)** | Yes: {scope, budget, ceiling, expiry, revocation}, [two-signer](./the-envelope.md) | No | No | No | No | No |
| **Machine payment** | Yes: 402 Offers; agent wallet (MPP / x402); [proof-before-account](./recipe-pay-a-402.md) | No | No | No | No | No |
| **Evaluation honesty** | GRADE: date-blinded, contamination-fenced, VS ungated / VS null; [a backtest is never flattering](./recipe-grade-a-strategy-honestly.md) | User-run backtests | Self-scored (bias-prone) | Framework backtests | N/A | Reported performance |
| **Provenance** | Certified Blotters + Grades + Proof URL | None standardized | Ad hoc | Logs | Audit trails (internal) | Statements |
| **Live model (custody)** | Never: BYO broker via OAuth, no custody | N/A | Your own broker | Your own broker | N/A | Vendor custodies |
| **Data licensing** | Databento served as derived works + BYO Alpaca; no raw redistribution | Vendor terms | You license raw feeds | You license raw feeds | Licensed, terminal-bound | Bundled |
| **Activation path** | [Proof-before-account](./recipe-pay-a-402.md): agent gets a trial capability, no card | Sign up | Install libs | Install / cloud sign-up | Sales + contract | KYC + funding |
| **Pricing** | Free sim/catalog; paid boundary as an HTTP 402 Offer | Freemium / subscription | Open-source / free | Open-source / hosted tiers | Enterprise seat license | AUM fee |
| **Best for** | An LLM agent perceiving and acting under bounded authority | A human watching charts | A human researching signals | A deterministic algo with no LLM | An analyst spanning many asset classes | A client who wants funds managed |

## The short version

Kestrel is not a general trading platform with an agent bolted on. It is the interface an LLM needs to perceive a market and act inside bounded authority, and nothing more. Everywhere that agent is missing, or where you want a vendor to hold the money, or where you just need raw bars or a chart to click, there is a better tool, and this page names it. That honesty is not a hedge; it is the point. [Sell the airport, not the pilots](./the-envelope.md), and tell people plainly when they do not need an airport.

> Kestrel is the wrong tool whenever a human is the primary reader, no LLM is in the loop, or you want a vendor to hold your money; it exists only for the agent that perceives in text and acts under an Envelope.