CATALYST/EDGE

Source-linked market intelligence for AI agents.

Catalyst Edge is a local, read-only Model Context Protocol server for public-company research. Ask an agent what changed for a ticker, why it matters, what contradicts it, and which sources support the answer.

It combines direct SEC filings and ownership records with optional, policy-gated sources into a compact evidence dossier. Every result preserves its source links and missing-data warnings.

Local first. Evidence and configuration stay on your machine. Research only. The deterministic scorer is untrained and unbacktested; Catalyst Edge does not provide investment advice, trading signals, or execution.

Install

Catalyst Edge requires Python 3.10+ and uv. The SEC requires an identifiable User-Agent; use your organization and a monitored email address.

Codex

codex mcp add catalyst-edge \
  --env 'CATALYST_EDGE_SEC_USER_AGENT=YOUR_ORGANIZATION YOUR_EMAIL' \
  --env 'CATALYST_EDGE_EVIDENCE_STORE=/absolute/local/path/evidence.sqlite3' \
  -- uvx --from 'catalyst-edge-mcp==0.1.8' catalyst-edge-mcp

Start a fresh task and verify that Codex discovers these two tools:

Claude Desktop

Download catalyst-edge-mcp-0.1.8.mcpb, then choose Settings → Extensions → Advanced settings → Install Extension…. Enter the same SEC identity when prompted. The extension is an unsigned custom bundle; review the source and published checksum before accepting Claude Desktop's warning.

Related MCP server: mentu-navigator-mcp

Use

Ask your agent a focused research question, for example:

What changed for NVDA in the last 14 days? Include sources, missing evidence, and anything that would weaken the conclusion.

The primary tool accepts a ticker, a 1–90 day lookback, source inclusion, and a research context:

{
  "ticker": "NVDA",
  "lookback_days": 14,
  "include_sources": true,
  "include_raw_signals": false,
  "risk_mode": "research"
}

risk_mode also supports alert_triage and thesis_review. Ticker validation runs before any provider is composed; invalid inputs fail clearly rather than producing a partial score.

From a terminal

# Run the local stdio MCP server
uvx --from 'catalyst-edge-mcp==0.1.8' catalyst-edge-mcp

# Get a dossier directly
uvx --from 'catalyst-edge-mcp==0.1.8' catalyst-edge-score NVDA --lookback-days 14

What it uses

The default evidence store is local SQLite at ~/.local/state/catalyst-edge-mcp/evidence.sqlite3. Set CATALYST_EDGE_EVIDENCE_STORE to choose another local path.

Check local readiness

CATALYST_EDGE_SEC_USER_AGENT='YOUR_ORGANIZATION YOUR_EMAIL' \
uvx --from 'catalyst-edge-mcp==0.1.8' catalyst-edge-smoke NVDA --lookback-days 14

The smoke check reports sanitized configuration, provenance, coverage, and readiness status. It never prints credentials or provider payloads.

How to read a result

Each dossier includes a deterministic score, direction, confidence, source-linked evidence, missing or stale families, and next checks. research.disposition tells an agent whether to review the evidence now, monitor it, or report insufficient evidence; it prioritizes research only and is not a trade signal. model_status is always not_trained in this release. A neutral or no-data result is a valid answer: missing evidence is uncertainty, not bearish evidence.

Evidence is compact by design. Use catalyst_edge_claim_sources with a claim ID to retrieve its paginated source records, including canonical URLs, timestamps, hashes, parsers, and policy decisions.

{
  "ticker": "NVDA",
  "edge": {"score": 62, "direction": "bullish", "confidence": 0.69, "scoring_method": "deterministic_v1", "model_status": "not_trained"},
  "research": {"disposition": "review_now", "primary_claim_id": "clm_aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "supporting_claim_ids": [], "contradicting_claim_ids": [], "blocking_gaps": [], "next_action": "Open SEC accession 0001045810-26-000001 and review the filed item text and exhibits."},
  "data_quality": {"coverage": "partial", "missing_families": [], "warnings": ["Deterministic v1 scoring is not backtested."]}
}

{
  "ticker": "NVDA",
  "edge": {"score": 50, "direction": "neutral", "confidence": 0, "scoring_method": "deterministic_v1", "model_status": "not_trained"},
  "research": {"disposition": "monitor", "primary_claim_id": null, "supporting_claim_ids": [], "contradicting_claim_ids": [], "blocking_gaps": ["options_flow"], "next_action": "Check whether a sector-wide event explains the observation."},
  "data_quality": {"coverage": "none", "missing_families": ["options_flow"], "warnings": ["options_flow provider yfinance is private diagnostic only; no production evidence or coverage credit was granted."]}
}

{
  "ticker": "NVDA",
  "edge": {"score": 50, "direction": "neutral", "confidence": 0, "scoring_method": "deterministic_v1", "model_status": "not_trained"},
  "research": {"disposition": "insufficient_evidence", "primary_claim_id": null, "supporting_claim_ids": [], "contradicting_claim_ids": [], "blocking_gaps": ["filings_news", "insider_trading", "options_flow", "social", "technical"], "next_action": "Retry with lookback_days=30 to check a wider filing window."},
  "data_quality": {"coverage": "none", "missing_families": ["filings_news", "insider_trading", "options_flow", "social", "technical"], "warnings": ["No live evidence adapters are configured."]}
}

Privacy

Results and SQLite evidence remain on your machine. Ticker and issuer queries may be sent directly to whichever public-source providers you enable. The SEC identity is sent only to sec.gov as its required request User-Agent.

Read the Catalyst Edge Privacy Policy.

Build from source

uv sync --frozen --extra dev
uv run --frozen pytest
uv run --frozen ruff check .
uv build --no-sources --out-dir dist

Default tests are offline and use sanitized fixtures. The release workflow tests Python 3.10 and 3.14, MCP contracts, a clean build, and the packaged artifact.

License

MIT