v1.7: concise setup with remembered choices, conclusions first, expanded daily history and full method-dependency checks. Fresh confirmation is still required. Upgrade details and data-source boundaries.
Question → sourced evidence → frozen method stances → Bull/Bear challenge → PM decision + saved audit
Historical UI recording (MP4) · Historical report artifact (SOX, Chinese)
The recording predates the current 26-seat candidate. It demonstrates an earlier interface only—not current timing, method fidelity, data accuracy, four-code-host validation, or ChatGPT Work validation.
One question. An inspectable case.
AlphaCouncil turns a ticker question into a reviewable research process. Parallel evidence workers gather public sources, selected investment-method seats interpret the same dated facts, bull and bear challenge the case, and a portfolio manager records the decision and its invalidation conditions. Missing inputs stay visible instead of being filled with guesses.
The same repository supports Codex, Claude Code, OpenCode, and Grok Build, plus a tool-only ChatGPT Work developer-mode gateway. It classifies companies, ETFs, and market indices before research so a basket is not analyzed as if it were an operating company.
Install in Codex
Prerequisite: Node.js 18 or newer. Install with these two shell commands:
codex plugin marketplace add Zhao73/alphacouncil-agent
codex plugin add alphacouncil-agent@alphacouncil
Plugins load when Codex starts. Fully quit and restart Codex, open a new session, then enter this in the Codex composer:
@alphacouncil-agent analyze AAPL
For ChatGPT Work developer mode, Claude Code, OpenCode, Grok Build, Windows, troubleshooting, and the optional global npm command, use the complete install guide.
Free first run
Check the keyless public-data layer before starting a council:
# Codex
@alphacouncil-agent AAPL news
# Claude Code, OpenCode, or Grok Build
/alpha AAPL news
This check starts no council workers and requires no data-vendor key. For bounded research in Codex, use @alphacouncil-agent AAPL quick; on the three slash-command hosts, use /alpha AAPL quick.
Choose the depth before it runs
AlphaCouncil shows the work plan first. Full research asks separately for method seats, evidence breadth, and depth; the user confirms them before workers start. Full tiers use 15 / 30 / 60 minute ceilings—never a hidden token or currency estimate.
| Run | Model-call structure | Time ceiling |
|---|
| Data check | Keyless tools only; no council workers and no additional model fan-out | Outside the council tiers |
| Quick research | 4 evidence workers in parallel → 1–4 method seats in parallel → Bull and Bear in parallel → PM | 10 minutes |
| Full — fast | 8 core or exactly 11 all-scope evidence workers start together; each selected method stance is frozen deterministically before one isolated explanation worker; 3 debate rounds → PM | 15 minutes |
| Full — normal | Same confirmed roster, frozen-stance sequence, 3 debate rounds, and PM, with a larger depth envelope | 30 minutes |
| Full — slow | Same confirmed roster and stages with the largest depth envelope | 60 minutes |
These are queue-to-terminal persistence ceilings, not measured completion times. They guarantee
an explicit terminal record even when work is incomplete; a successful live fast run within 15
minutes has not yet been demonstrated across the four hosts.
Fast keeps the full contract but uses an auditable stage-aware reasoning profile and one shared
lifecycle budget per primary/retry/repair chain; retries cannot silently double a seat's cap.
When an instrument classification and typed-fact coverage are already available, the selector also
shows an eight-family advisory method match derived from all 26 physical pack manifests. An
explicit objective and holding horizon additionally calibrate the match and separate directional,
non-voting risk, and context-only contributions; out_of_scope is never a negative vote. For a
one-year directional request, the PM maps sourced base-case total return to one published rating
rubric instead of counting conservative method seats; the server binds the frozen price/currency
and recomputes that return from a same-currency target plus income. An out_of_scope seat remains visible in
the method bench but is structurally absent from the PM rating path. It is only a prefill: every pack remains
selectable, no run starts without explicit confirmation, and a missing classification produces no
guessed default. The output represents AI-generated method simulations—not human experts,
independent models, or a promise of profit. See
method-panel recommendation and seat evidence.
Only the slow run with all methods and all evidence workers enables the additional verification path; the other full tiers do not claim that extra check.
What you gain
| Benefit | What it changes |
|---|
| A council, not one answer | Evidence specialists, method seats, opposing cases, and a PM expose where agreement comes from. |
| A stance before the story | Each selected full-run method stance is fixed from structured inputs before its isolated explanation is written. |
| Claims you can trace | Material report claims must point to source IDs; missing evidence remains a stated gap. |
| Disagreement that survives synthesis | Three cross-examination rounds and persisted minority or opposing reports keep the losing case available for review. |
| The right research path for the asset | Companies use issuer evidence; ETFs use dated holdings look-through; indices use aggregate methodology. The first data check is keyless. |
How the architecture differs
This compares workflow shapes, not named products. A particular tool may implement a different design.
| Review concern | Single-model reply or common shared-context flow | AlphaCouncil |
|---|
| Correlated errors | One shared context can carry an early mistake into every later step | Evidence seats and opposing paths run in isolated workers; they may still use the same provider or model and are not independent models |
| Position formation | The position can be composed together with its explanation | A structured stance is frozen before explanatory prose |
| Source trace | Traceability depends on the prompt and host | Every material claim is required to carry a source ID |
| Minority view | Dissent can be folded into the final summary | Minority and opposing reports remain surfaced as review artifacts |
What the seats are—and are not
The method-seat formulas are AI-authored reconstructions of published methods, pending human review. The named practitioners have not reviewed or endorsed these seats. They are not impersonations, independent models, or validated replicas. A stance is a structured argument to check against its inputs and sources—not a validated investment model.
Current source evidence boundary: 26 provisional method seats, 0 validated method models, 0/8
registered-and-completed canonical evaluation runs, and 0/4 live-host end-to-end runs. Passing source tests
does not change those zeros.
Disclaimer
AlphaCouncil is for education and research only. It is not investment advice, a recommendation, or a solicitation. AI-generated analysis can be incomplete, outdated, or wrong. Verify the evidence yourself and consult a licensed professional before making an investment decision. The authors accept no liability for losses.
Go deeper
Runtime outputs are written outside the repository under ~/.alphacouncil-agent/runs/<run_id>/.
Evidence first. Disagreement visible. Decisions reviewable.
↑ Back to top
