ALTA

Autonomous LLM Trading Asterism

A virtual trading platform operated by specialized LLM agents.

简体中文 · 繁體中文 · Quick start · Architecture · Verification · Website

Start with the opportunity. Choose the instrument that best expresses it.

ALTA is a local, multi-agent market-research system with an operator console, internal simulation and a separate broker execution boundary. Four specialist Scouts pursue leads; independent reviewers challenge the evidence; a strategy desk compares ways to act. Every handoff leaves an inspectable record.

The idea is a virtual research firm, not a chatbot that recommends a ticker: a thesis must survive disagreement, costs and time before it deserves capital.

[!IMPORTANT] Experimental research software—not investment advice, a production OMS or evidence of profitable Alpha. Execution has two modes: Shadow and Broker API. Broker orders require a verified, explicitly selected account and separate authorization. Six connectors exist; live-account acceptance is not established. Current execution limits ↓

What you get

ALTA brings three normally disconnected jobs into one local workspace. The output is not just a list of securities: it is a research trail you can challenge, an explicit decision, and a way to compare that decision with what happened next.

For researchers, this makes disagreement and source quality inspectable. For developers, it separates Agent judgment from scheduling, data contracts and execution. For an operator, it provides one place to see what ran, what remains uncertain and what—if anything—has permission to act.

See the work

Follow an opportunity from its source evidence to reviewer decisions. Open any image for the full view; saved hypotheses are not approved trades.

Research team	Inside an opportunity
Who investigated, which model ran and what it produced.	The thesis, supporting evidence and reviewer decisions.

Execution controls	Model routing
Two modes, an explicit destination and separate authority.	Choose models by role while preserving independent review.

Research captures: September 12, 2026; execution setup: September 13. English UI; Agent artifacts retain their authored language. The setup capture shows an unauthorized configuration, not a connected broker. Scope and image hashes: research review · execution review. No credentials or account details.

Run locally

Requirements: macOS or Linux, Node.js 22+ with npm, Python 3.12, uv, Docker/OrbStack and sufficient disk space. Initial setup needs internet; building the pinned custom harness also needs its Rust toolchain.

git clone https://github.com/kyky2347/ALTA.git
cd ALTA
npm run dashboard

The command installs locked frontend dependencies when needed, builds the console and opens an authenticated loopback page. Continue in the console:

Saving a key never grants trading permission. Model access, data entitlements and broker account permissions are separate requirements. A healthy service may be waiting for its next cycle; a Wait decision can be the correct result.

Setup and troubleshooting · Operator guide · Reproducibility

Read the console like a research desk

Start with a question, not a counter. The views share record IDs, so a discovery can be followed into the exact Run, review and event that changed its state.

The inspector exposes submitted summaries, structured artifacts and provenance, not a model's hidden chain-of-thought. Switching the interface language does not rewrite the underlying research record. Historical playback and the live edge remain distinct, so an old decision is not mistaken for new activity.

How the firm works

flowchart LR
    R["Investigate  
Scouts → evidence → thesis"] --> J["Challenge  
debate → rank → trade plan"]
    J --> G{"Independent audit  
and execution gates"}
    G -->|Ready| E["Act & observe  
execute → monitor → exit"]
    G -->|Not ready| W["Wait  
record the missing evidence"]
    W -. "focused follow-up" .-> R
    E -. "outcomes & Trader Mind updates" .-> R

Research is open-ended; authority is explicit. LLMs choose what to investigate and propose how to use it. Code enforces freshness, risk, account isolation and durable transitions. A persuasive thesis cannot approve its own execution.

Four minds, different questions

Each Scout has a Trader Mind—its research style, interests and accumulated lessons. They share tools, not a single agenda.

The toolset includes date-aware search, issuer crawling, feeds, public datasets and archive lookups. News and social claims are leads, not verified conclusions. Event time stays separate from retrieval time: a newly fetched page may describe an old event. Bounded budgets, source pacing and recorded partial results keep research inspectable. Retrieval design and limits →

Who owns each decision

flowchart TB
    O["One versioned opportunity  
same cited evidence"] --> A["Thesis assessor  
private, locked view"]
    O --> B["Independent challenger  
private, locked view"]
    A --> M["Moderator  
surface disagreements"]
    B --> M
    M --> R["Code-based ranking  
quality, uncertainty and edge gates"]
    R --> E["Expression Agent  
compare up to three payoff plans"]
    E --> U["Independent auditor  
select one plan or Wait"]

The assessors commit their views before seeing the other assessment. The moderator works from those saved records; it cannot replace a missing source with consensus. Ranking is software, not an extra Agent voting for its favorite. Market and portfolio checks sit between a proposed expression and its audit.

Only structured, versioned artifacts cross these handoffs—not an ever-growing shared chat. Models are configurable by role, with separate choices required for opposing assessments and for expression versus audit. Model diversity is a useful control, not proof that the reviewers are statistically independent.

From an idea to an accountable decision

Stocks, ETFs and supported option expressions can be research candidates; the broker executor has a narrower scope below. The obvious ticker may be illiquid, already repriced or duplicate another position's risk. Wait preserves the reason and next research question instead of forcing a trade.

The three-language console links evidence, Agent work, decisions and event history by record ID. Counts of saved records, completed work and currently running Agents are separate; none is a proxy for investment performance.

What travels with an opportunity

An Opportunity is the durable research identity; a Candidate is an incoming lead, and an Expression is a proposed way to act. Deduplication and later versions keep follow-ups attached to the original question instead of turning every wake into another apparently new idea.

Opportunity · stable identity, versioned thesis
├── Evidence     sources · event/observation times · content hashes
├── Hypothesis   causal claim · falsifier · expected horizon
├── Reviews      locked assessments · disagreement · decision
├── Expression   instrument · size · cost · independent audit
└── Follow-up    open questions · observations · exit/outcome records

These are linked records, not fields that an Agent can freely rewrite. A reader can trace what was known at the decision time without silently importing later knowledge. Several headlines about one event need not become several trades.

An opportunity can outlive a research cycle

The research agenda retains unanswered questions and catalyst deadlines. Later wakes can assign a focused follow-up while other Scouts continue exploring. Closing a tab does not cancel that agenda; the durable backend owns it. This is how the design accommodates a thesis that needs days or weeks to resolve, without keeping one chat session alive throughout.

Trader Mind experience and portfolio context guide attention, but are not promoted into source Evidence. Forward results are attributed to the frozen entry decision; feedback and forecast calibration have maturity gates. Existing calibration can reduce risk when evidence deteriorates, not autonomously relax the execution policy after a few lucky outcomes. Long-horizon reliability and investment skill still require sustained observation.

Two modes, one explicit destination

Configure → Verify account → Save destination → Authorize → Start
               Paper / Live is an account setting, not a third mode

Authorization binds the provider, account, environment and configuration revision. The backend accepts persisted research and independent audit artifacts—not browser-supplied orders. A separate monitor refreshes quotes, reconciles orders and manages exits without waiting for an LLM research cycle.

These are implementation states, not six live-account acceptance results. The initial broker lifecycle supports one active plan per account, long USD stocks/ETFs, whole shares and DAY limit orders. Options, shorts and multi-plan portfolio execution are outside that boundary. The legacy Tiger Paper executor remains isolated for compatibility, not a third selectable mode.

Use an empty, dedicated account for initial authorization. Revoke, reconcile and resolve existing exposure before switching. Exits are software-managed, not broker-native protective orders; downtime can delay them. Stopping ALTA does not close broker positions. Provider matrix and operating limits →

Under the hood

flowchart TB
    UI["React console · three languages"] <-->|"authenticated control & reads"| N["Node gateway  
launch · credentials · bounded tools"]
    N <--> P["Python research service  
scheduler · research lifecycle · API/SSE"]
    P <--> C["Codex App Server  
specialized LLM sessions"]
    P <--> DB[("PostgreSQL  
records · events · checkpoints")]
    P --> X{"Execution mode"}
    X --> S["Shadow  
internal positions & fills"]
    X --> B["Broker API  
selected account · separate authority · durable ledger"]

The browser is a control surface, not the scheduler. A managed backend owns research and recovery; PostgreSQL preserves its records, while Redis supports coordination. Broker intents enter isolated durable ledgers before submission. Uncertain orders reconcile by identity rather than being blindly retried.

alta-dashboard/                React console · English / 简体中文 / 繁體中文
alta-src/                      Node gateway · bounded tools · launch/control
alta-runtime/python/           research lifecycle · PostgreSQL · API/SSE
alta-runtime/capital-python/   isolated Tiger Paper executor
alta-runtime/broker-python/    isolated broker connectors and execution
vendor/openai-codex/            pinned, attributed Agent harness
docs/                          architecture · operations · verification

ALTA is useful for studying multi-agent judgment and building inspectable financial workflows. It is not a low-latency trading engine or a turnkey institutional stack. Recovery mechanisms are not a 24×7 uptime certification.

When reality interrupts

Research progress, browser connectivity and broker state are separate concerns. The system should not infer that an order failed merely because a response was lost, or infer that research stopped merely because a tab disconnected.

These are engineering mechanisms with test coverage, not a promise that every outage is recoverable unattended. Broker sessions, fresh data and a running backend remain necessary for software-managed exits. Unresolved positions or ambiguous orders may require operator review before work can safely continue.

Verify the work

The September 13 review records 916 passing tests across Node, research, legacy Paper, broker and frontend suites, plus lint/build checks, clean-source installation and sampled browser checks. No real account was authorized and no broker order was sent in that review. Earlier no-order research run →

Three kinds of evidence answer different questions:

A successful search is not a verified thesis; an approved thesis is not a fill; a profitable observation is not established Alpha. Keep those measurements separate when evaluating or extending the project.

Run the offline/contract checks without paid APIs:

./alta env setup --dev
./alta test
./alta env python -m pytest -q alta-runtime/python/tests
uv run --frozen --project alta-runtime/capital-python pytest -q alta-runtime/capital-python/tests
uv run --frozen --all-extras --project alta-runtime/broker-python pytest -q alta-runtime/broker-python/tests
node --test alta-dashboard/tests/*.test.mjs
corepack pnpm check
./alta env down

Integration tests need managed PostgreSQL; plain pytest without DATABASE_URL is not the complete check. See Reproducibility for the full gate, deterministic replay and clean-install scope.

Still unproven: sustained out-of-sample Alpha, live-account order acceptance, every power-loss scenario and universal browser compatibility. Test results describe tested behavior, not a guarantee of profit or fault-free operation.

Architecture · 24×7 runbook · Security · Contributing · Changelog

License and provenance

ALTA-authored source is Apache-2.0. A modified, pinned Apache-2.0 OpenAI Codex snapshot supplies the local App Server/harness; ALTA owns the research lifecycle and execution policy. Other packages and APIs retain their own licenses and terms. See ATTRIBUTION.md, NOTICE and the vendored change record.

Independently maintained; not endorsed by OpenAI or any named provider. Operators remain responsible for provider terms, exchange entitlements and data-redistribution rights.