An AI-powered multi-agent pipeline that reads insurance denial letters and writes back.

Counterclaim takes a denial letter, fans out to parallel evidence-gathering agents, synthesizes an appeal strategy, and produces a complete, case-specific appeal letter — all without a human touching a keyboard. It is also deployable as an ASI:One-compatible agent on Fetch.ai's Agentverse.

  [ Denial Letter / Upload ]
           │
           ▼
    App Backend (8000)          serves the React UI & manages async job queue
           │
           ▼
        Parser                  structures the raw denial document
           │
           ▼
  ┌────────┼────────┐
  │        │        │           ← parallel
Contact  Personal  External
 Agent   Evidence  Evidence
(8002)   (8003)    (8004)
  │        │        │
  └────────┬────────┘
           │
           ▼
  Appeal Strategy Agent (8005)  Claude: synthesizes denial + all evidence
           │
           ▼
    Drafting Agent (8006)       writes the final appeal letter
           │
           ▼
  [ Appeal Letter + Exhibits ]

Each agent is an independent FastAPI service. They communicate exclusively via JSON. The orchestrator wires them together end-to-end (via direct imports, not HTTP).

Counterclaim ships with an ASI:One-compatible coordinator at fetch_agents/counterclaim_coordinator.py. It implements the mandatory Agent Chat Protocol and exposes the full pipeline as a discoverable Agentverse workflow.

• ASI1 Chat Session: https://asi1.ai/shared-chat/93bd0964-153c-4a24-8061-a1265b9b5a9d
• Agent Profile: https://agentverse.ai/agents/details/agent1qwv5hzqy6vq4g8srs3d5hjzxjkss9p0tacvmhdkecus8fk55z5rdv6leevj/profile

Each agent manages its own virtual environment. Run these steps inside the agent folder you're working on:

git clone https://github.com/harshildalall/war-room.git
cd war-room/<your_agent_folder>
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

echo 'ANTHROPIC_API_KEY=sk-ant-your_key_here' > .env

Never commit .env. It is in .gitignore — keep it that way.

Start the app backend — it serves the React UI and orchestrates the pipeline:

cd app_backend
source .venv/bin/activate
uvicorn main:app --reload --port 8000

Then open http://localhost:8000 in your browser.

uvicorn main:app --reload --port <YOUR_PORT>

Use the port for your agent from the table above. The appeal_strategy_agent is the exception — run it with:

cd appeal_strategy_agent
uvicorn appeal_strategy.api:app --reload --port 8005

cd fetch_agents
echo 'FETCH_AGENT_SEED=your_seed_phrase_here' > .env
external_evidence_agent/.venv/bin/python fetch_agents/counterclaim_coordinator.py

Open the Agentverse Inspector URL printed in the terminal to register the agent. Once registered, it accepts direct chat prompts from ASI:One, e.g.:

Run the demo appeal case.

Every JSON message passed between agents must include:

{
  "case_id": "string",
  "schema_version": "string",
  "status": "string",
  "provenance": {}
}

The canonical source of truth for all schemas is shared/schemas.py. Do not modify it without announcing the change in the group chat first.

Every agent must expose:

• POST /run — accepts the agent's input JSON, returns its output JSON
• GET /health — returns {"status": "ok"}

Exception: appeal_strategy_agent uses POST /strategy and POST /strategy/validate.

• Never commit .env
• Never touch another person's agent folder
• Never modify shared/schemas.py without a group chat announcement
• Pass JSON between agents — never raw files
• All JSON output must include case_id, schema_version, status, provenance

git pull origin main        # always pull before starting
git add .
git commit -m "feat: your message"
git push origin main

The sample_inputs/ directory contains example JSON files for each agent stage. Use these to test your agent in isolation before plugging into the orchestrator.

• Python 3.11+
• FastAPI — agent HTTP servers
• Anthropic Claude API (claude-sonnet-4-6) — LLM backbone for parser, strategy, and drafting agents
• MongoDB — evidence store (counterclaim.evidence_chunks) and appeals history
• React 18 — frontend UI (served from app_backend)
• PyMuPDF — PDF text extraction (with regex fallback)
• uAgents / Fetch.ai — Agentverse deployment
• Pydantic — schema validation via shared/schemas.py