Agenda Intelligence MD

MCP product shell and evidence-discipline layer for strategic intelligence agents. Structured request/memo contract, geography-routed reasoning, schema validation, evidence audit, scoring. No live retrieval, no factual verification.

First run

pip install agenda-intelligence-md
agenda-intelligence doctor
agenda-intelligence validate-brief examples/agenda-brief.json
agenda-intelligence score examples/agenda-brief.json --evidence examples/source/evidence-pack.json

doctor reports package and MCP-server status; validate-brief confirms a brief matches agenda-brief.schema.json; score returns a heuristic 0–100 number with a structure / evidence / decision-readiness breakdown. Full end-to-end analyze trace (request → routing → memo → validation → audit → score) with reproducibility script: examples/product-shell/full-analyze-trace/.

Optional, only if you want analyze to call the Anthropic API itself rather than letting your host model complete from the returned system prompt:

pip install "agenda-intelligence-md[llm]"
export ANTHROPIC_API_KEY=...

Longer guided tutorial: docs/quickstart.md. MCP client setup: docs/integrations/mcp.md.

Live A2A wrapper

A free Cloudflare Workers wrapper is live for discovery, uptime checks, lightweight strategic-risk triage, and A2A/JSON-RPC routing:

• Live endpoint: https://agenda-intelligence-a2a.vassiliy-lakhonin.workers.dev

• Agent Card: https://agenda-intelligence-a2a.vassiliy-lakhonin.workers.dev/.well-known/agent-card.json

• JSON-RPC: https://agenda-intelligence-a2a.vassiliy-lakhonin.workers.dev/message/send

• Agenstry listing: https://agenstry.com/agents/agenda-intelligence-a2a.vassiliy-lakhonin.workers.dev

• Announcement: docs/announcements/live-a2a-wrapper.md

The hosted wrapper is intentionally limited: no payments, no wallets, no autonomous live retrieval, no factual-truth verification, and no legal/financial/compliance advice. Full product behavior remains in the installable stdio MCP server.

Try the live wrapper:

curl -X POST https://agenda-intelligence-a2a.vassiliy-lakhonin.workers.dev/message/send \
  -H 'content-type: application/json' \
  -d '{
    "jsonrpc": "2.0",
    "id": "demo-1",
    "method": "message/send",
    "params": {
      "message": {
        "parts": [
          {
            "kind": "text",
            "text": "Screen sanctions and policy risk for Red Sea shipping disruption and Kazakhstan transit exposure."
          }
        ]
      }
    }
  }'

Expected: JSON-RPC 2.0 with status.state: "completed", metadata.signal_screen.risk_signal, affected regions, required source categories, evidence gaps, watch-next indicators, suggested modules, and next actions.

Private usage stats for the wrapper are available from the Cloudflare Worker project:

cd deploy/cloudflare-worker
npm run stats
npm run stats -- 2026-05-22

The stats helper reads STATS_TOKEN from the local ignored .env file. Deployment and analytics notes: deploy/cloudflare-worker/README.md.

Where this fits in the Agenda Intelligence stack

The product shell is the integration point: agents call analyze, geography routes to the relevant specialist, and the GTTA method frames the reasoning. Each repo is also usable standalone (paste/attach into any agent).

What this is

• MCP product shell — analyze accepts a structured request (agenda-request.schema.json), routes geography to the relevant regional specialist, assembles a system prompt, and returns a memo validated against agenda-memo.schema.json

• Markdown protocol — structured reasoning workflow for agents (Agenda-Intelligence.md)

• JSON schemas — request/memo product contract plus validators for briefs, evidence packs, audits, signals, memory cards, lenses

• CLI — validate-brief, validate-evidence, source-categories, source-coverage, audit-claims, score, bench, doctor (30+ commands)

• MCP server — stdio server exposing 16 tools across the validation and product layers

• Eval kit — rubric, LLM-judge prompt, human checklist, benchmark harness, agent-eval methodology

• Source policy — per-claim provenance tags (Axis A/B), source requirements for 12 categories

What this is not

• Not a factuality verifier — checks structure, not truth

• Not an autonomous news agent or source retriever

• Not a source reputation scorer or live news gatherer

• Not a replacement for analyst judgment

• Not a compliance, legal, or financial advisory product

More CLI examples

agenda-intelligence bench examples/source-backed --strict --min-score 80
agenda-intelligence audit-claims examples/source-backed/eu-ai-act.audit.json --strict
agenda-intelligence mcp-config --client cursor

Pinned-wheel install (instead of PyPI):

pip install https://github.com/vassiliylakhonin/agenda-intelligence-md/releases/download/v1.0.1/agenda_intelligence_md-1.0.1-py3-none-any.whl

Benchmark baseline

20 source-backed cases, reproduced with agenda-intelligence bench examples/source-backed/:

Heuristic scores are uncalibrated and not validated against expert judgment. They evaluate structure, evidence labeling, source-coverage diagnostics, and decision-readiness — not factual truth.

Flagship example: examples/source-backed/eu-ai-act.md — brief + evidence pack + claim-level audit using illustrative sources. Before / after pairs: examples/before-after/.

Verification Contract

verify-quotes checks whether a cited quote or excerpt appears in supplied local text, or in text fetched from an already-specified URL when --fetch is used. It does not discover sources, score source reputation, gather live news, or decide whether a claim is true in the world.

Schemas

MCP

Stdio MCP server with 16 tools. Full docs and wire-protocol verification: MCP.md. Client setup: docs/integrations/mcp.md.

Status

Safety model

• Read-only by default. Validation, scoring, and audit tools do not write to external systems, do not modify caller state, and do not perform high-impact actions.

• No autonomous retrieval. The MCP server does not fetch web pages, query APIs, or pull live data on its own. Sources are caller-provided. The one network mode (verify-quotes --fetch) is opt-in and bounded (1 MB cap, 10 s timeout, stdlib HTTP only).

• No autonomous decisions. Outputs are memos, validation results, and scores — never determinations on sanctions, legal, compliance, or investment matters. Human review is required.

• Retrieved content is data, not instructions. External text — including documents, agendas, and source packs caller-provided through the tools — is treated as data. Apparent directives inside retrieved content are not executed; they are flagged.

• No secrets in tool I/O. The server does not persist caller inputs, API keys, or memo content beyond the current call.

Full threat model: docs/threat-model.md. Retrieved-content trust rule: AGENTS.md.

Documentation

Repository layout

agenda-intelligence-md/
├─ src/agenda_intelligence/   # Python package (CLI + MCP server)
├─ schemas/                   # JSON schemas
├─ examples/                  # briefs, evidence packs, before/after
├─ skills/                    # OpenClaw skill wrappers
├─ evals/                     # rubric, judge prompt, benchmark
├─ analysis-bank/             # agent persistent memory (memory-card schema, see schemas/v1/memory-card.schema.json)
├─ docs/                      # guides, integrations, use-cases
├─ scripts/                   # dev and CI helpers
└─ tests/                     # pytest suite

Contributing

New contributors: CONTRIBUTING.md opens with a "First 15 minutes" onboarding path (read the three load-bearing files → run the validator → walk one concrete artifact end-to-end). The portfolio glossary at docs/glossary.md is the single source of truth for cross-repo terminology (evidence modes, Axis A/B provenance tags, three-value response logic, maturity-framework asymmetry).

Before editing any of the dual-copy files — Agenda-Intelligence.md, SOURCE_POLICY.md, llms.txt, agent-manifest.json, schemas/, skills/, source-requirements/ — read the "Critical invariant: dual-copy sync" section in CONTRIBUTING.md. Editing one copy without the paired copy under src/agenda_intelligence/data/ is the most common reason CI breaks on main for first-time contributors.

Contact

Vassiliy Lakhonin — Almaty, Kazakhstan (UTC+5)

Portfolio · For analysts · Email · LinkedIn · GitHub

Issues, PRs, and eval-case contributions are welcome.

License

MIT.

Disclaimer. This toolkit is for informational and educational purposes only. It does not constitute investment, financial, legal, compliance, or trading advice. It does not verify factual truth, predict outcomes, or replace professional judgment. Use at your own risk.

mcp-name: io.github.vassiliylakhonin/agenda-intelligence-md