agentii-investment-intelligence

agentii-investment-intelligence

The financial data layer for AI agents.
Open-source alternative to FactSet / Daloopa / S&P Global for AI agents.
1,146+ US equities with full SEC filing history. 80 skills across 14 verticals. 30 MCP tools.
One API key. Zero infrastructure. Single entrance: /agentii:skill-name.

[!WARNING] Not investment advice. This repository is software. It produces analyst work product for review by a qualified professional — never a recommendation to trade. Read the full Disclaimer before using anything here.

Why agentii

Wall Street pays $30K+/seat/year for FactSet, Bloomberg, and S&P Global. Those platforms were built for humans clicking through terminals. AI agents need agent-use-ready data — structured, citation-backed, page-addressable, API-delivered.

agentii.ai ingests every SEC filing (10-K, 10-Q, 8-K, 20-F, 6-K) and 15K+ earnings call transcripts (2022+) into a Neon PostgreSQL data plane with 15.99M XBRL facts, 51K+ source documents, and 1.34M+ parsed pages. A single agentii MCP server at mcp.agentii.ai exposes 30 tools (incl. institutional ownership + insider activity) that Claude Code, OpenCode, Goose, Codex, OpenClaw, and Claude Cowork consume natively.

This repository mirrors anthropics/financial-services — marketplace plugin system, vertical skill decomposition, agent-plugin bundling. The difference: all skills point at a single agentii MCP server backed by agentii.ai’s own data plane. There is no second MCP — office output uses the same code-mode approach (Python + LibreOffice, invoked via Bash) that Anthropic’s package uses.

Quick Install

1. Get an API Key

agentii.ai/api-keys — 7-day free trial, 2,000 credits, no credit card.

2. Global MCP Setup

export AGENTII_API_KEY=sk_live_YOUR_KEY_HERE
claude mcp add-json --scope user agentii \
  '{"type":"http","url":"https://mcp.agentii.ai/mcp","headers":{"Authorization":"Bearer <YOUR_KEY>"}}'

Writes to ~/.claude.json. Restart Claude Code — all 30 tools auto-discover on every session from any directory.

3. Install Skills (Claude Code: Local Copy)

bash scripts/copy-skills-local.sh ~   # ~ = the target; skills land in ~/.claude/

The target argument is not optional in practice. The script defaults to the current directory, so running it bare from the repo root installs into <repo>/.claude/ instead of ~/.claude/. Pass ~ unless you specifically want a project-local install.

Restart Claude Code — skills register under a single unified namespace, /agentii:skill-name (skills land in ~/.claude/skills/agentii/, commands in ~/.claude/commands/agentii/). This is the recommended install method on Claude Code — one namespace, no per-vertical prefixes, and it works reliably on all Claude Code versions.

The installer is idempotent and prunes: re-running it removes any installed skill that no longer exists upstream (a renamed or split skill, for example), and prints each one by name. Pass --dry-run to preview without writing.

Single namespace by design: the local-copy path exposes only /agentii:*. There is no /equity-research-core:* or /models-and-pitches:* surface — every skill is reached the same way regardless of which vertical authored it.

Other hosts: copy-skills-local.sh writes to ~/.claude/ only. For OpenCode, Codex, Goose, OpenClaw and Cowork, see For Other CLI Agents — the skills are portable markdown, but the install path differs per host.

claude plugin marketplace add agentii-ai/agentii-investment-intelligence
claude plugin install models-and-pitches     # adds /models-and-pitches:* namespace
# ... etc for other verticals

4. Verify

/agentii:recent-quarter LLY

Expected: structured, citation-backed report with real SEC filing data and clickable citations like [📄 LLY 10-K p.42](https://agentii.ai/v/LLY/sec129/42) — every material fact is immediately followed by its source link. The closing TUI reply includes a Key Citations block of clickable URLs so you can cmd+click straight to the exact SEC page.

What’s Inside

Skills

80 skills across 14 verticals. Each is a skills/agentii/<name>/SKILL.md file with YAML frontmatter and markdown methodology — the single canonical artifact across all hosts. Thin commands/*.md wrappers ship for explicit /agentii:skill-name slash-command invocation.

Start here

All valuation skills support --mode=scenario for Bear/Base/Bull probability-weighted analysis.

Browse all 80: the full registry with descriptions, modes and tool allow-lists is skill-registry.yaml. Methodology depth lives in per-skill references/ directories (progressive disclosure) — the SKILL.md body stays lean (~700–900 words) and detail loads on demand. Full methodology →

Governance & Research Modes

Spec 046 added the layer that makes the research checkable rather than merely generated. It is the largest recent body of work and it ships in the scenarios vertical.

Two operating modes

agentii.md has two roles, and the filesystem decides which. With a constitution.md present it is a chronicle (a memory index, rotated per period). With none, it is the constitution — and rotating it would rotate away the project’s principles. The detector reads the filesystem; it keeps no new state.

The constitution

/agentii:constitution scaffolds a workspace’s investment constitution: constitution.md (prose + SemVer amendment log), constitution.yaml (executable position/concentration constraints, budgets, regime drift triggers), plus assumptions.yaml, value-checks.yaml and taxonomy.yaml. Theses compile against a constitution_pin, so a stale pin is a hard failure rather than a silent drift.

Gates, not vibes

Correctness is enforced by dispatch preconditions and write-boundary gates, not by the order in which an agent chooses to work:

• G1 — deterministic: pure script, millisecond, zero LLM. Citation integrity, numeric canonical form, VACUOUS reporting, evidence class.
• G2 — judgement: an independent-context validator sub-agent at phase boundaries.
• G3 — human audit: only the G1/G2 red items, via theses/{nnn}-{slug}/checklists/*.md.

A gate that did not run must say so (mechanism_outcome: VACUOUS) — an un-run gate that reports success is the failure mode the whole layer exists to prevent.

Report pipeline

pack → author → assemble → render. Skills produce artifacts; the packer assembles them into a report input; the author writes content.html against a fixed outline; the assembler injects page furniture — headers, footers, page numbers, table of contents, and the disclaimer page — and the render step iterates on the rendered output rather than on the source.

Office Output (Code-Mode + LibreOffice)

v3.3.0 continues Anthropic’s proven code-mode architecture — no office MCP server. The agent writes self-contained Python scripts and executes them via Bash. Full contract →

Conventions (mirroring Anthropic xlsx-author): blue font = hardcoded input, black = formula, green = cross-sheet link. A Checks tab carries TRUE/FALSE validation ties. LibreOffice headless (soffice --headless) handles recalculation and PDF export. The formulas-over-hardcodes invariant (hardcoded_count == 0 for projection/discount/PV cells) is mandatory per FR-020.

Layered preflight (FR-043): skills probe for a live Office session (Cowork mcp__office__* tools) first, then fall back to Python + LibreOffice, then degrade to .md with the exact pip install remediation command. Never a silent failure.

Workspace Memory

A file-first hybrid memory architecture that persists context across sessions. After running skills, your workspace looks like this:

workspace/
├── constitution.md                     # Thesis-mode governance (spec 046); agentii.md where absent
├── agentii.md                          # Memory index — or THE constitution, if no constitution.md
├── style.md                            # Optional workspace overrides (currency, peers, verbosity)
├── NVDA/
│   ├── 2026-06-15_0930_recent-quarter_summary.md   # per-skill outputs, YAML frontmatter
│   └── 2026-06-15_1045_dcf_base.xlsx               # Office artifacts
├── snapshots/
│   └── NVDA/
│       └── 2026-06-15_thesis.md        # Point-in-time synthesis, restored on session start
├── sessions/
│   ├── INDEX.md                         # Session index (auto-loaded)
│   └── 2026-06-15/                     # Full transcripts (on-demand only)
├── _cross/                              # Multi-ticker analyses (peer-bench, comps, competitive-positioning)
│   └── semis_2026-06-15_1400_peer-bench_nvda-amd-avgo.md
└── _sector/                             # Pure sector/thematic analyses (names lowercase-hyphenated)
    └── tech-semiconductors/
        └── 2026-06-15_1500_sector-overview_summary.md

Key conventions:

• agentii.md — one file, two roles, decided by whether constitution.md exists: a chronicle (append-only memory index, rotated per period) or the constitution itself (principles, never rotated). Machine-parseable via head -20. Appended after every skill run; entries are never modified or deleted. Schema →
• {ticker}/ — per-skill outputs. Frontmatter carries a public core required in both modes: as_of (the date the analysis is as-of), skill, affix, key_metrics, conclusions, claims[], mechanism_outcome, plus ticker or tickers. Schema →
• snapshots/{ticker}/{YYYY-MM-DD}_{semantic-slug}.md — point-in-time synthesis (≤400 words) that states which prior conclusions are confirmed, updated, or superseded. The key is the ticker, not the thesis id, because a ticker always exists and a thesis id does not — thesis attribution lives in frontmatter, never in the path. Contract →
• Claim classification is a field, not a badge. Every claim carries claim_class ∈ {FACT, DEDUCTED, VIEW}; the inline [FACT]/[DEDUCTED]/[VIEW] badges are rendering from it, and the field wins on disagreement. facts_count / deducted_count / views_count are derived from that list, never authored. The reason is that a gate which counts claims by reading prose has stopped being deterministic — which is the one thing a G1 gate must be.
• Artifacts go to artifacts/ in thesis mode only. In single-skill mode there is no artifacts/ to hold them, and agentii.md + style.md + snapshots/ + sessions/ keep their original behaviour.
• _cross/ — multi-ticker outputs. _sector/ — industry/thematic outputs with no primary ticker.
• sessions/ — transcripts archived by date (not auto-loaded); INDEX.md is auto-loaded. Format →

Citations: Page-Accurate Provenance

Every material fact, table row, and metric in a deliverable is immediately followed by its clickable source link — not deferred to a bottom appendix. Inline-first placement is the package’s core UVP.

Revenue grew 22% YoY to $215.9B [📄 NVDA 10-K p.42](https://agentii.ai/v/NVDA/sec173/42)

The bottom ## Citations section provides a non-duplicative roll-up index. The closing TUI reply includes a Key Citations block — the headline 5–10 facts as clickable URLs, so you can cmd+click straight to the exact SEC page without opening the deliverable file.

Citation format: https://agentii.ai/v/{ticker}/{citation_id}/{N} — path-based, ~7 tokens, browser-redirects to the exact filing page. Earnings-call transcripts use the ect<N> id form.

Data-Source Priority

Every skill follows a mandatory data-source ordering (FR-075):

1. XBRL facts FIRST (grounding truth) — search_xbrl_facts with view=detailed for segment/product/channel breakdowns
2. SEC filings SECOND — 10-K (annual), 10-Q (quarterly), 20-F/6-K (foreign) via the three-layer retrieval protocol
3. Web search LAST RESORT — only when both XBRL and SEC filings are insufficient; flagged web_search_used: true in frontmatter; non-authoritative

Three-Layer Retrieval Protocol

Skills that search unstructured documents at scale follow a mandatory protocol codified in contracts/retrieval.md:

search_cross_period is the primary multi-period path for skills analyzing 4+ fiscal quarters.

Coverage

1,146+ US public companies across med + tech + industrial + fin + consumer sectors, with full SEC filing history from 2022 onward (10-K, 10-Q, 8-K, 20-F, 6-K) and earnings call transcripts (2022+). Every data point carries a clickable citation watermark linking to the original filing page.

Data volume: 15.99M XBRL facts, 51K+ source documents (SEC filings + earnings call transcripts), 1.34M+ parsed silver pages. XBRL facts updated daily via Dagster pipeline. SEC filings indexed within hours of EDGAR publication. Full coverage → | Request a ticker →

Skills surface a data_freshness warning for tickers with < 100% coverage and refuse to fabricate data outside the launch cohort.

Architecture

┌─────────────────┐     ┌──────────────────┐     ┌────────────────────┐
│  AI Agent        │     │  MCP Server      │     │  REST API          │
│  (Claude Code,   │ ──► │  mcp.agentii.ai  │ ──► │  api.agentii.ai    │
│   OpenCode, etc) │     │  30 tools        │     │  Hono + Vercel     │
└─────────────────┘     └──────────────────┘     └────────┬───────────┘
                                                          │
                          ┌───────────────────────────────┤
                          │                               │
                    ┌─────▼──────┐                  ┌─────▼──────┐
                    │  Neon      │                  │  Redis     │
                    │  PostgreSQL│                  │  (Upstash) │
                    │  15.99M    │                  │  tracing   │
                    │  XBRL facts│                  │  hot tier  │
                    └────────────┘                  └────────────┘

Data plane: Neon PostgreSQL (product data — XBRL facts, companies, filings, entity aliases). Tracing plane: Redis Upstash (hot, 7d TTL) + Supabase PostgreSQL (cold, audit). Office: code-mode Python + LibreOffice — no office MCP server.

One MCP server. One API key. Zero infrastructure.

Pricing

7-day free trial, 2,000 credits, no credit card required. Early adopter pricing — your rate stays as coverage grows. Full pricing →

For Other CLI Agents

The skills are portable markdown following the open Agent Skills standard. What differs per host is the install path, and not every host has an adapter shipped in this repository:

# Recommended: install the full agentii namespace
cp -r plugins/agentii-plugin/skills/agentii ~/.claude/skills/agentii/    # Claude Code
cp -r plugins/agentii-plugin/skills/agentii ~/.config/opencode/skills/   # OpenCode
cp -r plugins/agentii-plugin/skills/agentii ~/.codex/skills/             # Codex
cp -r plugins/agentii-plugin/skills/agentii ~/.config/goose/skills/      # Goose
openclaw add ./plugins/agentii-plugin                                      # OpenClaw

# Or: single vertical for lightweight installs
cp -r plugins/vertical-plugins/equity-research-core/skills/agentii ~/.config/opencode/skills/

Two MCP transports. The quick-start above uses the hosted HTTP server at https://mcp.agentii.ai/mcp. The files under adapters/ configure the stdio package (npx -y @agentii/investment-intelligence) instead. Pick one deliberately — they are not the same process, and mixing them produces a confusing “tools not found”.

See adapters/ for per-host configuration files, and docs/install/ for step-by-step guides. All agents benefit from ai-agents.txt at the repo root.

Making It Yours

• Bring your templates — mount firm-branded .pptx templates at ./templates/ for pitch-deck and earnings-preview
• Adjust methodology — edit ## Defaults tables and references/institutional-defaults.md
• Override via style.md — per-workspace style.md overrides defaults for lookback quarters, reporting currency, peer universe, and output verbosity
• Set your own limits — /agentii:constitution writes the position caps, concentration limits and regime drift triggers a workspace is governed by
• Chain skills — dcf → pitch-deck for end-to-end model-to-deck workflows; xlsx-financials → audit-xls for quality assurance
• Edit skills in plugins/vertical-plugins/<vertical>/skills/agentii/<name>/SKILL.md — the single canonical source
• Sync changes: python3 scripts/sync-agent-skills.py then bash scripts/assemble-agentii-namespace.sh
• Run CI checks: python3 scripts/check.py before pushing — validates manifests, frontmatter, CI gates, and cross-file consistency

Troubleshooting

Repository Structure

agentii-investment-intelligence/
├── plugins/
│   ├── agentii-plugin/                  # Meta-plugin: /agentii:* surface (70 symlinked skills)
│   ├── vertical-plugins/                # 14 verticals, 80 skills
│   │   ├── equity-research-core/        # 9   company dimensions
│   │   ├── models-and-pitches/          # 9   models + decks
│   │   ├── bio-pharm/                   # 15  FDA catalysts, trials
│   │   ├── scenarios/                   # 10  spec-046 research kit
│   │   ├── quantitative-analysis/       # 5
│   │   ├── idea-generation/             # 5
│   │   ├── options-derivatives/         # 5
│   │   ├── business-intelligence/       # 4
│   │   ├── industry-analysis/           # 4
│   │   ├── macro-strategy/              # 4
│   │   ├── portfolio-strategy/          # 4
│   │   ├── technical-analysis/          # 4
│   │   ├── risk-and-psychology/         # 1
│   │   └── trading-as-business/         # 1
│   └── agent-plugins/
│       └── agentii-equity-agent/        # Managed agent bundle
├── contracts/                           # 34 shared contracts (single source of truth)
├── data-tools/                          # Instant macro/market/earnings data (spec 039)
├── scripts/                             # CI gates, sync, validation, assembly, report pipeline
├── packaging/                           # 4 host export targets (spec 039)
├── docs/                                # install guides, architecture, CLI surfaces
├── adapters/                            # Per-host MCP config (5 hosts)
├── workflows/                           # Enrichment presets (spec 039)
├── style.md                             # Package-shipped formatting standard
├── skill-registry.yaml                  # 80 skills — the registry
├── QUICKSTART.md                        # Step-by-step walkthrough
├── README.md, LICENSE, NOTICE, CHANGELOG.md
└── SKILL.md                             # Root package manifest

Contributing

Everything is markdown, YAML, and Python. Fork, edit, PR.

• Read QUICKSTART.md first for a working install, and browse docs/ for the architecture notes.
• Edit skills in plugins/vertical-plugins/<vertical>/skills/agentii/<name>/SKILL.md — the single canonical source
• Sync changes: python3 scripts/sync-agent-skills.py then bash scripts/assemble-agentii-namespace.sh
• Run python3 scripts/check.py before pushing — validates all manifests, frontmatter, CI gates, and cross-file consistency
• Skills follow the open Agent Skills standard, supported by Claude Code, OpenCode, Codex, OpenClaw, Goose, and Claude Cowork

Disclaimer

This repository is software. It does not provide investment, legal, tax, or accounting advice, and nothing it produces is a recommendation, an offer, or a solicitation to buy or sell any security.

The skills generate analyst work product for review by a qualified professional. Outputs are staged for human sign-off and are not intended to be acted on unreviewed. Any figure, valuation, model, or conclusion a skill produces may be incomplete, delayed, or wrong — it is derived from the sources cited inline, those sources may themselves be wrong, and the analysis may have misread them. No representation or warranty is made as to accuracy or completeness.

Statements about the future are forward-looking and inherently uncertain. Past performance is not indicative of future results. You are responsible for your own due diligence and should consult your own advisers before acting on anything produced with this software. The authors and distributors accept no liability for any loss arising from reliance on it.

Data is provided by agentii.ai and third-party sources under their own terms. Market data may be delayed. Coverage is not universal — skills surface a data_freshness warning and are designed to refuse rather than fabricate, but absence of a warning is not a guarantee of completeness.

Generated research outputs carry their own disclaimer. Every presentation-shaped output — thesis-report.html, dashboard.html, pitch-deck, earnings-preview — must include the canonical block from plugins/vertical-plugins/scenarios/templates/disclaimer.md, which is the single authored source. It must be included verbatim, with placeholders filled; it is never restated, paraphrased, or forked, and scripts/check_disclaimer.py fails the build if it drifts.

See LICENSE for the software licence and NOTICE for attribution.

License

Apache License 2.0 © agentii-ai. See LICENSE and NOTICE.

This package includes skills ported and optimized from anthropics/financial-services (Apache 2.0). Methodology bodies stay byte-stable from upstream; data-source blocks and tool calls are replaced with agentii-native equivalents.