Argus — Enterprise AI Intelligence & Research Platform

Argus turns a research question into a verified, cited report. It plans the research, collects evidence from the web and from an organisation's own documents, analyses it with AI agents, checks every claim against its sources, flags contradictions, and keeps watching the web for changes that matter - for many organisations at once, each isolated from every other.

Security is the design constraint, not a feature added at the end: tenant isolation is enforced by PostgreSQL row-level security, everything fetched from the web or uploaded is treated as hostile, AI agents can only use tools they were explicitly given, and every security-relevant action lands in a tamper-evident audit log.

What it does

Area	Highlights
Research jobs	plan → collect → analyse → verify → detect contradictions → report; budgets, human approvals, progress streaming; reports in Markdown, JSON, CSV and PDF
Knowledge base	PDF, DOCX, HTML, Markdown, CSV, JSON and text; malware scan, parsing in a sandboxed subprocess, chunking, embeddings; hybrid keyword + vector search with authorisation inside the SQL
Secure web research	SSRF guard with DNS pinning and per-hop redirect checks, size/time/decompression limits, robots.txt and per-domain politeness, source reputation
AI agents and LLM gateway	declared tool catalogue, per-agent permissions and budgets, kill switches; Claude (Anthropic), OpenAI-compatible and offline local models; data-classification policy before any text leaves
Evidence verification	every finding needs a verified quote; figures must appear in the evidence; contradictions and confidence scores; prompt-injection defences end to end
Continuous monitoring	watched pages and searches, change significance scoring, in-app and e-mail notifications
SaaS operation	organisations, roles, restricted projects, API keys and service accounts; plans and quotas; retention; suspension, deletion with a grace period, restore, purge; owner data export
Security center	audit chain verified daily, security events and recommendations, denied-access auditing, kill-switch administration
Observability	OpenTelemetry traces, Prometheus metrics, alert rules with runbooks, Grafana dashboard
Web dashboard	/app: projects, active jobs, reports, sources, knowledge base, monitoring, alerts, AI usage and costs, system health, security events

Architecture

            browser (/app)  ·  API clients (bearer tokens, API keys)
                              │ HTTPS
                ┌─────────────▼──────────────┐
                │ argus api (FastAPI)        │  authentication, RBAC, rate limits, CSP, audit
                └─────────────┬──────────────┘
  modules: identity · tenancy · research · sources · documents · knowledge · llm · agents ·
           monitoring · notifications · security_center · evaluation · platform
                              │
   PostgreSQL 16 + pgvector (RLS)  ·  Redis (limits, cache)  ·  object storage (encrypted)
                              ▲
   argus worker (documents, research, monitoring)  ·  argus scheduler (periodic duties)

One installable package (src/argus): a modular monolith whose layers (apps → modules → security → infrastructure → core) are enforced by import-linter. Details: system architecture · database design · AI architecture · security model · threat model.

Quick start

Option A - Docker Compose

cp .env.example .env                        # then replace every change-me password
uv run argus keys generate --quote >> .env  # signing, encryption and audit keys
docker compose up --build                   # PostgreSQL, Redis, Mailpit, migrate, api, worker, scheduler
docker compose run --rm api users create --email you@example.com --name "Your Name"

Open http://localhost:8000/ (the dashboard) and http://localhost:8000/docs (interactive API docs, development only). Real malware scanning: docker compose --profile scanning up --build with ARGUS_DOCUMENTS__SCANNER=clamav.

Option B - local processes, no Docker

uv sync --group localdb                                       # Python 3.12 environment + pgserver
uv run python scripts/local_postgres.py start --data var/pg   # prints the DSNs to put in .env
uv run argus keys generate --quote >> .env
uv run argus db migrate                                       # as the schema owner role
uv run argus serve                                            # terminal 1: http://localhost:8000
uv run argus worker                                           # terminal 2
uv run argus scheduler                                        # terminal 3
uv run python scripts/seed_demo.py                            # optional: a demo organisation

seed_demo.py creates a demo organisation with documents, a real report (from the offline models), a queued job and a monitor change, and writes the demo sign-in to .env.demo (git-ignored).

No API key is needed to try Argus: without provider keys it runs on deterministic offline models. To use Claude, install the extra (uv sync --extra anthropic) and set ARGUS_LLM__ANTHROPIC_API_KEY. For Claude Opus 5.5 and Sonnet 5.5 the server-side refusal fallback is enabled by default; turn it off with ARGUS_LLM__ANTHROPIC_REFUSAL_FALLBACK=false.

Configuration

Every setting is an environment variable ARGUS_<SECTION>__<NAME> (a .env file is read too); .env.example lists the important ones. argus config check validates the configuration for its environment: staging and production refuse unsafe values (public API docs, wildcard hosts or origins, non-https URLs, missing secrets, unencrypted database connections, the development scanner) and name the variable to fix. argus config show prints the effective configuration with secrets redacted.

Setting	Default	Purpose
ARGUS_ENVIRONMENT	development	development, testing, staging or production
ARGUS_DATABASE__URL / ARGUS_DATABASE__MIGRATION_URL	-	runtime role / schema owner (migrations and operator commands only)
ARGUS_REDIS__URL	none	shared rate limits and caches; in-process fallbacks in development
ARGUS_PLATFORM__DEFAULT_PLAN	enterprise	plan of new organisations (free for a hosted service)
ARGUS_HTTP__WEB_DASHBOARD	true	serve the dashboard at /app
ARGUS_LLM__ANTHROPIC_API_KEY	none	Claude; without any provider key the offline models are used
ARGUS_OBSERVABILITY__OTEL_ENABLED	false	export traces over OTLP

Command line

Command	What it does
argus serve · worker · scheduler	the three process roles
argus db migrate · current · check	migrations (owner role); drift check between models and migrations
argus keys generate	new key material for .env or the secret store
argus users create · disable · enable	administrative accounts; containing a compromised account in one step
argus orgs list · suspend · resume · restore · set-plan · purge	organisation lifecycle and plans (audited, with a reason)
argus audit verify · export · prune	verify every chain, write an evidence copy, retention behind a signed checkpoint
argus killswitch	stop an agent, tool, provider, model or all AI, platform-wide
argus jobs · prompts · eval · config	queue inspection, prompt deployments, evaluation gate, configuration

Tests and quality gates

uv run pytest                     # unit and security tests (database tests are skipped)
ARGUS_TEST_DATABASE_URL=postgresql://postgres@127.0.0.1:5432/postgres uv run pytest   # everything: + integration, API surface, red team, evaluation
uv run pytest -m perf             # benchmarks with latency budgets (needs the database)
uv run ruff check . && uv run ruff format --check . && uv run mypy && uv run lint-imports
uv run bandit -c pyproject.toml -r src

Integration tests create a throwaway database per run and run as the runtime role, so row-level security takes part in every test. CI adds dependency auditing, CodeQL, secret scanning, container scanning, signed releases with SBOM and provenance, and Kubernetes manifest validation.

Documentation

• Roadmap and a teaching guide per phase in docs/phases: purpose, architecture, security, files, tests, common mistakes, acceptance criteria.
• Operations: deployment · backup and restore · incident response · disaster recovery · alerts.
• Decisions: architecture decision records and the review of the original specification.
• Release history: changelog.

Status

Phases 1-25 of the roadmap are complete, except phase 18 (knowledge graph), which was left out by decision. Version 0.1.0.

Security

Please report vulnerabilities privately - see the security policy. The design is documented in the security model and threat model.

Contributing

Contributions are welcome: read the contributing guide and the code of conduct first.

License

MIT