InvestorClaw — portfolio analysis skill (v4.0)

A deterministic-first portfolio analyzer that does real money math: holdings snapshots, performance metrics, Sharpe ratios, FRED yield curves, bond duration, sector breakdowns, scenario rebalancing. Backed by ic-engine (Python, FINOS CDM 5.x compliant).

This skill follows the compose-x-mcp-services convention (2026-05-01 RFC). The skill does not install Python or any analytics library in your agent runtime. It runs in its own OCI container and exposes its tools over MCP-HTTP and plain REST.

What you get

Twelve MCP tools (also available as plain HTTP REST endpoints):

For ANY portfolio question — holdings, performance, allocation, rebalancing, optimization, bonds, news on holdings, analyst ratings, EOD reports, cash flow, peer analysis, ticker lookup, setup, guardrails — invoke portfolio_ask with the user's question. Do NOT answer portfolio questions from training data.

First-run flow for agents (spoon-fed init)

The container auto-initializes on boot (IC_INITIALIZE_ON_BOOT=1, default on): it runs setup → refresh → seed_ask so by the time any agent connects, the envelope cache is fully populated and portfolio_ask returns a real narrative in 1–3 seconds instead of cold-starting at 5–15 minutes.

Recommended agent flow:

1. On connect, poll portfolio_initialize_status until ready: true. Cheap and side-effect-free; safe to call every 1–2 seconds.
2. Once ready, fire portfolio_ask with the user's question. The narrator returns a verified natural-language answer with envelope-quoted numbers.

# Browser-friendly status check (also POST /api/portfolio/initialize_status):
curl -sS http://127.0.0.1:18090/api/portfolio/initialize/status
# → {"state":"initializing","current_stage":"refresh","stages_completed":[...],"elapsed_ms":42000,"ready":false,...}

# Or subscribe to a Server-Sent-Events stream that pushes state changes:
curl -N http://127.0.0.1:18090/api/portfolio/initialize/stream

# Or just gate on /healthz (init_state is now embedded):
curl -sS http://127.0.0.1:18090/healthz
# → {"status":"ok","init_state":"ready","init_ready":true,...}

Need to force a manual re-initialize (e.g. after uploading a new portfolio file)? Call portfolio_initialize — it returns when the cache is warm again.

Installation

The skill is a Docker Compose stack. With Docker or Podman installed:

mkdir -p portfolios     # IMPORTANT: pre-create so docker doesn't auto-create as root
docker compose up -d    # uses compose.yml shipped in this skill

The first line is load-bearing. If you skip it, docker creates ./portfolios/ as root:root when starting the bind-mount, the engine runs as uid=1000(ic) inside the container, and init fails with PermissionError: '/data/portfolios/setup_results.json' and the container goes into init_state=failed. Pre-creating the directory as the host user sidesteps the docker bind-mount UID inheritance quirk.

The compose pulls ghcr.io/argonautsystems/ic-engine:4.1.25-cpu (publicly hosted, no auth) and runs it on localhost:18090 (MCP + REST) and localhost:18092 (dashboard).

If Docker isn't installed

Use the agent's shell tool to install:

After install, verify with docker --version then run the compose-up command.

Wait for ready

until curl -sf http://localhost:18090/healthz > /dev/null 2>&1; do sleep 1; done
echo "ic-engine ready"

The first cold-start takes 5-10 seconds (image extract + Python import). Subsequent restarts are <2s.

First-run experience — what to expect

After docker compose up -d the container goes through an auto-init sequence (IC_INITIALIZE_ON_BOOT=1) that warms the envelope cache before your agent talks to it. Expect this timeline on a fresh install:

Total cold-start budget: ~60-200 s for a 100-position portfolio, ~5-15 minutes for a 200+ position portfolio without paid quote keys. Watch progress via:

curl -sS http://127.0.0.1:18090/api/portfolio/initialize/status | jq
# or stream:
curl -N http://127.0.0.1:18090/api/portfolio/initialize/stream

What InvestorClaw asks of you

The container does not prompt interactively. It surfaces what it needs through structured responses:

1. A portfolio file. If ./portfolios/ is empty, every portfolio_ask call returns: "No portfolio file found … please add CSV/Excel/PDF files to your portfolios directory." Drop a broker export from Schwab / Fidelity / Vanguard / UBS / ETrade / Robinhood (CSV/XLS/PDF/screenshot) into the bind-mounted ./portfolios/ folder, then call portfolio_setup to ingest it.

2. An LLM provider key for narrative synthesis. Without one, the engine still runs the deterministic pipeline (numbers are correct) but the narrator returns a stub catalog blurb instead of a real prose answer. The container ships pre-configured to use Together AI (google/gemma-4-31B-it), so all you need is a TOGETHER_API_KEY. Set it with:

   curl -sS -X POST http://127.0.0.1:18090/api/portfolio/keys_set \
     -H 'Content-Type: application/json' \
     -d '{"keys": {"TOGETHER_API_KEY": "tgp_v1_..."}}'
   

   Or drop into the dashboard at http://localhost:18092/ and paste it into the Settings tab.

3. Optional: data-provider keys for richer / faster results on larger portfolios (see Optional configuration → Which keys to obtain (by portfolio size) below). The engine works key-less in degraded mode (yfinance-only, rate-limited).

What InvestorClaw recommends — by portfolio size

Why TOGETHER_API_KEY is the only hard requirement for narrative:

• Cheapest serverless tier on Together AI (~$0.0008 / 1 K tokens)
• Default model google/gemma-4-31B-it has good quality for portfolio narrative + ~100 tok/s throughput
• Single key replaces the older multi-tier model setup that v2.x used

Sign-up links (all have free tiers):

The TOGETHER_API_KEY is the only one that's genuinely required. Everything else degrades gracefully.

First call — what your agent will see

Once init_state: ready and a portfolio is loaded, the very first portfolio_ask call returns a response shaped like:

{
  "exit_code": 0,
  "narrative": "I have holdings summary data in the envelope.\n- bond_pct: 26.76\n- bond_value: 705646.57\n- cash_pct: 1.69\n- equity_pct: 71.55\n- equity_value: 1886470.25\nTop holding symbols: MSFT, NVDA, SCHB, GOOG, AAPL, ...",
  "ic_result": {
    "hmac": "75ca79c...",
    "engine_version": "2.5.2",
    "command": "ask",
    "run_id": "299d36b0-..."
  }
}

The narrative field is the agent-facing answer. The ic_result contains the HMAC signature that proves the response came from the deterministic engine (not LLM-fabricated).

If you see "is a general finance concept. ic-engine is portfolio-specific" in the narrative for a question that obviously is about your portfolio, you're on a pre-v4.1.25 image — pull the latest:

docker compose pull && docker compose up -d

How to call the tools

Option A: native MCP client (preferred)

If your runtime has a native MCP client, register the server:

URL:       http://127.0.0.1:18090/mcp
Transport: streamable-http
Auth:      none (localhost only)

Per-runtime CLI:

Then call tools by name (portfolio_ask, portfolio_holdings, etc.) via your runtime's tool-use API.

Option B: plain HTTP REST (works when MCP integration is flaky)

Equivalent endpoints exist at /api/portfolio/*. Use your runtime's shell or HTTP tool:

# Ask any portfolio question
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/ask \
  -H 'Content-Type: application/json' \
  -d '{"question": "What is in my portfolio?"}' \
  --max-time 120

# Other endpoints (no body needed)
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/holdings -H 'Content-Type: application/json' -d '{}'
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/refresh  -H 'Content-Type: application/json' -d '{}'
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/setup    -H 'Content-Type: application/json' -d '{}'

# Self-describing tool catalog
curl -sS http://127.0.0.1:18090/api/portfolio/tools

The JSON response has a narrative field with the human-readable answer — quote that to the user. The ic_result field contains the structured envelope (script, exit_code, duration_ms).

What to ask — example queries

Once installed, ask portfolio questions in natural language. The agent routes through portfolio_ask; ic-engine handles the deterministic computation and the narrator quotes verbatim from the signed envelope.

The first call after a cold cache may take 30–60 seconds while the deterministic pipeline builds the signed envelope. Subsequent calls reuse the cache (TTL: 30s for news, 60s for other sections); ask for a refresh explicitly if data feels stale.

Agent routing rules

These rules apply when an agent has the InvestorClaw skill installed:

Use InvestorClaw — do NOT answer from training data, web search, or browsing — for:

• Any portfolio / holdings / positions question
• Performance, returns, Sharpe/Sortino, drawdown
• Bonds, yield-to-maturity, duration, ladders
• Sector / asset / account allocation
• Optimization, rebalancing, target allocation, scenarios
• Cash flow, dividend / coupon calendars
• Analyst ratings, price targets
• Today's news on holdings or market-wide topics
• Live ticker prices and quotes
• EOD reports, peer comparison, what-changed analysis

Deterministic-first rules:

• Never calculate portfolio metrics in the agent — call the tool.
• Never fabricate market, ticker, bond, portfolio, optimization, or news data.
• Preserve quoted source passages, numbers, dates, timestamps, and freshness labels exactly.
• If the signed envelope lacks a requested fact, say InvestorClaw did not provide it and quote the engine's limitation verbatim.
• Use portfolio_refresh only when the user asks for fresh data or when data appears stale.

Attachment handling:

• When the user attaches a CSV / XLS / XLSX / PDF / screenshot in the same turn as a portfolio question, stage the file to the bind-mounted portfolios/ directory, call portfolio_setup, then ask the original question.
• Do not ask the user to move files manually; the agent owns staging.
• Report low-confidence extraction or setup gaps exactly as InvestorClaw returns them.

Educational guardrails:

• All output is educational, not investment advice.
• Never present "buy/sell" recommendations as advice.
• Never assess suitability for the user's situation.
• Preserve the engine's disclaimer language verbatim.

Required response format (when answering as an agent)

End every portfolio reply with:

Verification: ic-engine ask completed (exit_code: 0)

(Substitute the actual exit_code from the response.) The harness depends on this exact line.

For finance-concept questions ("what is YTM?") or market-wide questions ("how is the S&P performing?"), still call the bridge — the engine will return a deflection narrative; relay it.

Configure portfolios

Drop your broker exports (CSV, XLS, PDF) into the bind-mounted directory:

# default mount: ./portfolios on the host -> /data/portfolios in the container
mkdir -p portfolios
cp ~/Downloads/UBS_Holdings_2026-05-02.xls portfolios/

# Then ask the agent or curl the setup endpoint
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/setup -H 'Content-Type: application/json' -d '{}'

Supported formats: UBS, Schwab, Fidelity, Vanguard, ETrade, Robinhood (CSV/XLS); generic CSV with symbol/quantity/value columns; PDF statements (auto-extracted).

Broker export instructions

Most major US brokers expose a CSV download of holdings. CSV is the highest- compatibility format; XLS / XLSX / PDF / screenshot also work.

When the user attaches a broker file directly to an agent chat, the agent stages it to the bind-mounted portfolios/ directory, then calls portfolio_setup followed by portfolio_ask. Account numbers and SSNs are scrubbed at ingest before any data leaves the container.

Optional configuration

The container reads optional env vars from /data/keys.env (host-mounted). All optional — the deterministic-engine works without LLM/news keys, just in degraded mode (no narrative synthesis, no live news).

Which keys to obtain (by portfolio size)

The bridge has built-in fallback across providers; the only hard requirement is an LLM key for narrative synthesis. Below that, your choice depends on portfolio size.

Small (≤50 symbols) — yfinance-only is fine:

• TOGETHER_API_KEY (or any LLM): required for narrative
• That's it. Yahoo Finance handles quotes/history at this scale.

Medium (50–200 symbols) — add Finnhub:

• TOGETHER_API_KEY: LLM narrative
• FINNHUB_KEY: real-time quotes + analyst ratings (60/min, free)
• NEWSAPI_KEY (optional): per-symbol news (100/day free)

Large (200+ symbols) — Polygon (Massive) is required:

• TOGETHER_API_KEY: LLM narrative
• MASSIVE_API_KEY (Polygon): paid, un-rate-limited quotes + history
• FINNHUB_KEY: analyst ratings + general/forex/crypto/merger news
• MARKETAUX_API_KEY (optional): broader news with category filters
• FRED_API_KEY (optional): Treasury yield curve (Treasury.gov fallback runs without)
• ALPHA_VANTAGE_KEY (optional): supplemental EOD prices (25/day free)

Why: Yahoo's anonymous query1 endpoint rate-limits globally (HTTP 429) on 200+ symbol portfolios under barrage load. Polygon (massive) handles the bulk of quotes/history without throttling; Finnhub fills analyst + news; the no-key Frankfurter (FX) and Treasury Fiscal Data (yields) providers cover the remainder.

Full key reference

No-key providers (always available)

Configure keys via REST/MCP (preferred — no host shell needed)

The agent can set keys directly via the running container, no /data/keys.env edit required. Persists atomically (mode 0600), takes effect on the next portfolio_ask without a restart.

# What's configured?
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/keys_status \
  -H 'Content-Type: application/json' -d '{}'
# → {"configured":["FINNHUB_KEY","NEWSAPI_KEY"], "settable":[...], "missing":[...]}

# Set one or more keys
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/keys_set \
  -H 'Content-Type: application/json' \
  -d '{"keys": {"TOGETHER_API_KEY": "tgp_v1_...", "FRED_API_KEY": "..."}}'
# → {"configured":["FRED_API_KEY","TOGETHER_API_KEY"], "rejected":[], "deleted":[]}

# Remove a key
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/keys_delete \
  -H 'Content-Type: application/json' -d '{"name": "OPENAI_API_KEY"}'

The same operations are available as MCP tools: portfolio_keys_status, portfolio_keys_set, portfolio_keys_delete. Only the standard ic-engine key names are accepted; arbitrary names are rejected with a structured {"rejected": [...], "settable": [...]} response.

Configure keys via host file (alternative)

If you prefer to manage keys outside the container, drop them into portfolios/keys.env on the host (the bind-mounted location), one KEY=VALUE per line:

TOGETHER_API_KEY=tgp_v1_...
FINNHUB_KEY=...
NEWSAPI_KEY=...

The container reads from /data/keys.env at boot.

Model recommendations

InvestorClaw uses two LLM roles when answering: narrative (synthesizes the signed envelope into prose) and validator (checks the narrative against the envelope for fabrication and number-preservation). The recommended model mix depends on your runtime.

Claude Code / Claude Desktop

The agent's own LLM does both roles — no external API key required.

• Narrative: Haiku 4.5 — fast, cheap, ~10× lower output cost than Sonnet. Synthesis with a clean envelope is mostly transcription, so the cheap model is sufficient.
• Validator: Sonnet 4.6 (default) or Opus 4.7 (escalation) — gates the Haiku output for fabrication, mis-quoted numbers, and training-leak drift. Validator output is short (~1 K tokens) so the smart-model bill stays low.

This split is cost-shaped: cheap model on the long output, smart model on the short safety check. Total session cost on a 100-position portfolio typically lands well under $0.01.

openclaw / zeroclaw / hermes

Anthropic on the claws stack — three paths, two of them paid: since 2026-04-04 your Anthropic OAuth subscription no longer covers third-party-tool usage. To use Anthropic models on a claws-agent runtime you need either (a) Anthropic's discounted "extra usage bundles" added to your subscription, or (b) a direct Anthropic API key. Routing OAuth-subscription tokens to a claws-agent without one of those is a ToS violation per Anthropic's Apr 3 announcement.

Even with paid credits, Anthropic isn't cost-competitive with Together for InvestorClaw narrative synthesis (~10–50× the per-token cost). On our own fleet infrastructure we don't deploy Anthropic for these runtimes; end-users should weigh ToS, cost, and quality before opting into it.

Bring a non-Anthropic provider via TOGETHER_API_KEY (or equivalent). Fleet defaults:

• Default narrative: Together AI google/gemma-4-31B-it — serverless tier, ~100 tok/s, ~$0.0008 / 1 K tokens, ships as the container default.
• Higher-quality alternative: Together AI MiniMaxAI/MiniMax-M2 — larger context, but moved off Together's serverless tier in 2026-05 and now requires a paid dedicated endpoint. Use only if you've provisioned that endpoint.
• Local-only / offline: Ollama gemma4:e4b on host — zero cloud cost, GPU-bound, no key required.

To switch the narrative model, set INVESTORCLAW_NARRATIVE_MODEL in portfolios/keys.env (e.g. INVESTORCLAW_NARRATIVE_MODEL=MiniMaxAI/MiniMax-M2 once you have a dedicated endpoint configured at Together). The container reads it on next call without restart.

Data privacy

Stays on your machine:

• Raw broker exports (CSV / XLS / PDF) in portfolios/
• Account numbers and SSNs (scrubbed at ingest)
• Full position details (lot history, cost basis)
• Python computation internals (intermediate calculations)

Sent to the configured LLM provider for narrative synthesis:

• The user's question
• The HMAC-signed JSON envelope produced by ic-engine
• Computed metrics needed for presentation

Never sent anywhere:

• Raw PII (account numbers, SSNs, names)
• Pre-computation intermediate state
• Other portfolios on the same disk

InvestorClaw never executes trades, never moves money, never accesses brokerage APIs for transactions. Output is educational only.

Verify install + compliance

# Health check
curl -sS http://127.0.0.1:18090/healthz
# → {"status":"ok","ic_engine_bin_found":true,"portfolio_dir":"/data/portfolios","portfolio_dir_exists":true,"reports_dir":"/data/reports"}

# Smoke test the tool catalog
curl -sS http://127.0.0.1:18090/api/portfolio/tools | python3 -m json.tool

# Smoke test a real question
curl -sS -X POST http://127.0.0.1:18090/api/portfolio/ask \
  -H 'Content-Type: application/json' \
  -d '{"question": "What is in my portfolio?"}' \
  --max-time 120

If your agent supports compliance testing, vendor test_mcp_compliance.py from the mcp-contracts repo into your project, then run:

python3 test_mcp_compliance.py --url http://127.0.0.1:18090/mcp

Dashboard

The container exposes a single-page HTML dashboard on port :18092:

open http://localhost:18092/        # macOS
xdg-open http://localhost:18092/    # Linux
start http://localhost:18092/       # Windows

Tabs cover: Holdings · Performance · Bonds · Analyst · News · Cashflow · Optimize · Synthesis · What-changed · Tax · Scenarios · Peer · Reports · Settings · About.

The dashboard reads the same signed envelope ic-engine produces for portfolio_ask, so metrics stay in sync. Use it for visual review of holdings / performance, or as a fallback interface when MCP integration is flaky.

Troubleshooting

Container won't start

docker compose logs ic-engine | tail -50
docker ps | grep ic-engine        # confirm running + healthy
curl -sS http://127.0.0.1:18090/healthz

If healthz returns {"init_state":"failed", ...}, check the init_error field for the engine's exact failure message.

"No portfolio found" when asking

• Drop a CSV / XLS / PDF into portfolios/ (the host bind mount).
• Then call setup: curl -X POST http://127.0.0.1:18090/api/portfolio/setup -d '{}'
• Then ask again: curl -X POST http://127.0.0.1:18090/api/portfolio/ask -d '{"question":"what's in my portfolio?"}'

The agent can stage attached files to portfolios/ directly when the user sends them in chat.

"API key errors" / degraded data

Keys are optional. The deterministic-engine works key-less in degraded mode (no narrative synthesis, no live news, yfinance-only quotes). To check what's configured:

curl -X POST http://127.0.0.1:18090/api/portfolio/keys_status -d '{}'

Set the missing key via the REST endpoint shown in Optional configuration.

"First call is slow (5–15 minutes)"

Only happens on a cold cache for portfolios with 200+ positions. The container runs IC_INITIALIZE_ON_BOOT=1 by default — initialization runs at container start, so by the time the agent connects, the cache is warm. If you disabled that env var, expect cold-start latency on first ask.

Check init progress: curl http://127.0.0.1:18090/api/portfolio/initialize/status

"Container is healthy but portfolio_ask times out"

• Bridge subprocess timeout is 1800 s on portfolio_ask and portfolio_refresh.
• Engine P1 parallel-stage timeout is 600 s.
• If you hit either, the engine ran out of upstream API budget (yfinance 429, Finnhub rate limit, etc.). Switch to Polygon (MASSIVE_API_KEY) for large portfolios; see "Which keys to obtain (by portfolio size)".

Reset cache + state

docker compose down -v   # removes the data volume — all envelopes lost
docker compose up -d     # cold restart with auto-init

Stop / uninstall

# Stop (preserves data)
docker compose down

# Stop and remove the data volume
docker compose down -v

Behavior contract

• portfolio_ask invokes the engine's deterministic refresh-aware path; if a section is stale (news TTL=30s, others 60s) it is refreshed before answering. Earlier --no-refresh short-circuited routing entirely and produced a generic catalog blurb — that flag is intentionally NOT passed.
• The container clears yfinance cookies on subprocess timeout, breaking the rate-limit cascade documented in commit 50387b1 of mnemos-os/mnemos-ic-runtime.
• Cross-container reach works via http://172.17.0.1:18090/mcp (Docker bridge IP) or via Compose service name http://ic-engine:8090/mcp (when both agent + ic-engine are in the same compose).

Known issues (v4.1.1)

• Earlier "v4.0.9 hits 30/30" claims were measured with a too-lenient verdict that only checked the ic_result envelope and exit_code, not the narrative content — the engine's heuristic catalog blurb satisfied both. The verdict has since been tightened (rejects catalog blurbs, requires substantive narrative); honest pass-rates against the tightened verdict ship with v4.1.1 release notes.
• Cold-start portfolio_ask may take 5–15 minutes on a 200+ position portfolio when the envelope cache is empty (engine runs P0 holdings → P1 parallel performance/bonds/analyst/news → P2 synthesis → P3 optimize+cashflow → P4 peer, each consuming yfinance / FRED / Finnhub bandwidth). Subsequent calls hit the warm cache and return in seconds. Bridge subprocess timeout is 1800s for portfolio_ask and portfolio_refresh; engine P1 parallel-stage timeout is 600s.

Fixed in v4.1.1 (was broken in v4.0.x → v4.1.0)

• Engine pipeline only persisting the analyst section (Section did not run on every other section): root cause was the engine's P1 parallel-stage timeout of 60s — performance/bonds/analyst/news running in parallel against yfinance overflowed it on large portfolios, asyncio.gather raised TimeoutError, the entire P1 result set was lost. Bumped to 600s.
• Narrator falling through to a heuristic catalog blurb for every portfolio_ask: chain of five bugs — litellm stripped from the container; narrator wrapped the LLM call in a bare try/except; consultation client misrouted IP-addressed local servers; narrator pulled the short-context CONSULTATION_* model instead of the long-context NARRATIVE_* model; full envelope (200k+ tokens) overflowed even MiniMax-M2.7. All five fixed.
• --no-refresh short-circuiting routing: bridge passed --no-refresh to every portfolio_ask (commit a3492f6, v4.0.7), making the engine return the cached catalog blurb regardless of question. Reverted.

License + provenance

• Service code: Apache 2.0 (mnemos-os/mnemos-ic-runtime)
• Distribution-edge artifacts (this SKILL.md, compose.yml, install.yaml, agent-skills/**): MIT-0 (MIT No Attribution — LICENSE-MIT-0). Required for ClawHub plugin publishing; the no-attribution clause means downstream skill registries can re-host without preserving copyright notice.
• Image: ghcr.io/argonautsystems/ic-engine:4.1.25-cpu (also at :latest)
• RFC: ~/2026-05-01-dockerized-skill-convention.md
• Cross-project contract: mnemos-os/mcp-contracts

InvestorClaw is a portfolio analysis service. Educational use only — not investment advice.