ALTA Research Trading Workbench
Use specialist agents to investigate market opportunities with traceable evidence, then simulate or execute only under explicit authorization.
- Source repo
- kyky2347/ALTA
- Stars
- ★ 1.1k
- Last updated
- 26d ago
- License
- Apache-2.0
- Primary language
- Python
- FA score
- 83/100 · Good
At a glance
- How it runs
- Works with
- Portable with changesCodex (Partial support)
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- A researcher tracking a filing, business change, or catalyst can use the Change / event Scout to create a sourced, timestamped opportunity.
- Not a fit if
- Trading teams that need verified live-account acceptance
- Institutions requiring low-latency execution or certified 24/7 uptime
- Users seeking investment advice or proof of profitability
- Source review
- 83/100 · Good
What does this agent do, and when should you use it?
ALTA is a local multi-agent market research system built around a React console, a Node gateway, a Python research service, and PostgreSQL records. Four Scouts investigate leads, independent reviewers challenge the evidence, and a strategy desk compares expressions; the resulting opportunity retains its sources, decisions, and follow-up history. Operators can manage internal positions in Shadow mode or configure a separate Broker API execution path, where orders require a verified, explicitly selected account and separate authorization. The documented setup requires macOS or Linux, Node.js 22+, Python 3.12, uv, Docker/OrbStack, and an internet connection for initial setup. ALTA is research software, not investment advice, a low-latency trading engine, or evidence of profitability.
Run npm run dashboard to install locked frontend dependencies when needed, build the React console, and open an authenticated loopback page. In the console, operators enter research or data credentials, assign models by role, and start managed research services. The Python research service schedules four Scout roles, which use date-aware search, issuer crawling, feeds, public datasets, and archive lookups to produce cited opportunity records. Independent reviewers submit locked assessments; a Moderator surfaces disagreements; software ranks opportunities and applies gates; an Expression Agent compares up to three plans; and an independent auditor selects a plan or Wait. ALTA stores evidence, decisions, events, simulated positions, monitoring, and outcome records. Shadow sends no broker orders; the separate Broker API path accepts persisted research and audit artifacts, while an independent monitor refreshes quotes, reconciles orders, and manages exits.
- A researcher tracking a filing, business change, or catalyst can use the Change / event Scout to create a sourced, timestamped opportunity.
- A quantitative research team examining price dislocations can use the Market dislocation Scout to gather leads and independent reviewers to test counterevidence.
- Researchers studying policy or industry spillovers can use the Causal / policy Scout to trace possible second-order effects across businesses.
- An investment research team comparing candidate stocks, ETFs, or supported options expressions and their costs can use the strategy desk to record plans and a possible Wait decision.
- A developer inspecting multi-agent workflows can use the console to review Agent roles, models, outputs, tool activity, context usage, and latency.
- An operator who wants to track positions without sending broker orders can use Shadow mode to record simulated fills, costs, positions, and exits.
How do you install or deploy this agent?
Requirements are 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 a Rust toolchain. Clone the repository and start the console:
git clone https://github.com/kyky2347/ALTA.git
cd ALTA
npm run dashboardFor first use, enter research or data credentials in the console, assign available models, and prepare the locked backend environment. Saving a key does not grant trading permission; model access, data entitlements, and broker account permissions are separate.
How do you use this agent?
In the console, connect research or data credentials, assign models under Agents → Agent models, prepare the backend environment, and start research services; begin in Shadow. Open a Scout record to inspect its citations and reviewer decisions. Use Research radar for opportunity details, Agent desk for work records, and Activity & history for events and replay. If configuring Broker API, verify the account, save an explicit Paper or Live destination, and authorize separately; only supported account and execution scopes are available. Stopping ALTA does not liquidate broker holdings.
What are this agent's strengths and limitations?
- Durable linked records connect opportunities, evidence, reviews, expressions, and outcomes, preserving what was known at decision time.
- Reviewers lock independent assessments before seeing one another's views, while software handles ranking and execution gates.
- Shadow mode manages internal fills, costs, positions, and exits without sending broker orders.
- The three-language console exposes event history and Agent activity for research and workflow inspection.
- Broker execution is separated from research and uses durable ledgers and identity-based reconciliation for uncertain responses.
- Deployment spans several runtimes and managed database/cache services, making setup substantial.
- Research requires model access, data credentials, and relevant entitlements; the README does not state service or API prices.
- Six connector implementations do not amount to six live-account acceptance results; Longbridge / Longport and Schwab authorization are blocked.
- The initial Broker API lifecycle is narrow: one active plan per account, long USD stocks/ETFs, whole shares, and DAY limit orders; options, shorts, and multi-plan execution are out of scope.
- Software-managed exits depend on a running backend and broker session; stopping ALTA does not close positions, and outages can delay exits.
- Sustained out-of-sample Alpha, live-account order acceptance, every power-loss scenario, and universal browser compatibility remain unproven.
How does this agent compare with similar options?
Compared with a chatbot that simply names a ticker, ALTA preserves leads, evidence, independent reviews, proposed expressions, and follow-up as linked records. Compared with direct broker execution, Shadow manages only internal simulated fills and positions; Broker API is a separate path requiring explicit account selection and authorization.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| ALTA Research Trading Workbench This agent | 83 · Good | Web app | ★ 1.1k | 26d ago | Python | — |
| QuantDinger AI Trading OS | 89 · Good | Self-hosted serviceFree + model costs | ★ 13k | 3d ago | Python | Codex · Claude Code · OpenAI API |
| TradingAgents for A-Shares | 52 · Major gaps | Library / SDKFree + model costs | ★ 3.7k | 2d ago | Python | OpenAI API · Claude API |
| Agentic RAG Survey Guide | Insufficient evidence | Library / SDK | ★ 1.7k | 11mo ago | — | — |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md, the README, and the supplied engine test material describe research-agent secret isolation, least privilege, explicit account authorization, durable order intents, retry blocking, and audit boundaries. Least privilege, user confirmation, data-flow transparency, sensitive-data handling, and external-effects controls score 3. Dependency security scores 2: lockfiles and CI audit gates are described, but no audit result for this review is supplied. Rollback scores 2 because revocation can enter close-only recovery, yet stopping ALTA does not liquidate holdings and exits depend on software monitoring. Source attribution scores 2: the Codex snapshot, NOTICE, and change record are identified, but publisher identity is unverified.
The documentation and test material give consistent design descriptions for account isolation, recovery, ambiguous order outcomes, and failure handling. Static evidence cannot establish runtime state, sustained availability, or complete dependency reachability, so self-consistency and dependency availability score 2. Error boundaries and tests against secret echo are described, but the supplied material does not establish complete operator guidance for every failure path; failure messages score 2.
The README addresses researchers, developers, and operators and explains research, Shadow, and controlled broker workflows; audience and scenarios and capability boundaries score 3. Trigger descriptions focus on console steps and runtime cycles without fully specifying what events trigger each capability, so trigger precision scores 2. Setup requirements are clear but limited to macOS/Linux, Node 22+, Python 3.12, uv, Docker/OrbStack, and internet for initial setup; environment fit scores 2.
The README gives clear entry points for architecture, operations, and troubleshooting, so information architecture scores 3. Local startup, requirements, and subsequent steps are concrete, so install notes score 3. Opportunity, Candidate, and Expression are defined, but multiple services and execution boundaries still add conceptual load; naming stability scores 2. Operating guidance and commands are supplied, but there is no dedicated FAQ, so examples and FAQ score 2. Risks, unproven areas, and operating limits are specific, so known limitations score 3. The repository includes an Apache-2.0 license, scoring 3. A CHANGELOG and revision metadata are referenced, but the supplied material does not show detailed version history, so versioning and changelog score 2. SECURITY.md gives a vulnerability-reporting path and the README says the project is independently maintained, but the responsible maintainer is not verified; maintenance responsibility scores 2.
The console links evidence, reviews, decisions, and event history while distinguishing research records from investment results; the supplied tests also describe offline contract and lifecycle cases. Output usability and marginal value score 3. The system requires model and data services, a database, and local infrastructure, while long-running operation can still require operator attention; cost-benefit scores 2.
The README links important claims to architecture, audit, reproducibility, and operations material, and the supplied tests contain concrete boundary assertions; claim traceability scores 3. The README, SECURITY.md, CI, and tests cross-support key boundaries, but test and audit outcomes remain claims in the supplied material and cannot be independently confirmed in this static review; cross-source corroboration scores 2. The materials distinguish research claims, test coverage, account acceptance, and investment performance and identify what remains unproven; fact-inference separation scores 3.
- This is a static review of the supplied files with low confidence; no code was run and test or audit results were not independently verified.
- The materials state that real-account acceptance is unproven. Broker exits are software-managed, downtime can delay them, and stopping ALTA does not close positions.
- Live authorization and real-account order acceptance remain unestablished; the stated initial boundary is one account and plan, USD stocks/ETFs, whole shares, and DAY limit orders.