A swarm intelligence engine that rehearses the future.

Feed it a document. Describe a scenario. Watch hundreds of AI agents with distinct personalities, memories, and social instincts interact — and return with a prediction.

_{One-click self-host — four services, one API key, ~60 seconds. Full walkthrough ↓}

---

Contents

• What It Does
• How It Works
• Quick Start — get an API key, install in 2 min
• Self-Host (Docker / Railway)
• MCP Server — tools available
• What's Different — vs the original MiroFish
• Persona Fidelity — the drift problem and how we fix it
• Monorepo Structure
• Use Cases
• Acknowledgments
• License

---

What It Does

DeepMiro extracts entities and relationships from any document — a policy draft, a market report, a chapter of a novel — and constructs a parallel digital world. Inside it, hundreds of autonomous agents form opinions, argue on simulated social platforms, shift allegiances, and produce emergent behavior that no single prompt could predict.

You get back a structured prediction report and a living world you can interrogate, agent by agent.

Input: A PDF and a question in plain language. Output: A detailed prediction report + an interactive simulation you can explore.

How It Works

Document ──► Entity Extraction ──► Agent Generation ──► Dual-Platform Simulation ──► Prediction Report
              (NER + GraphRAG)    (personas, memory,     (Twitter-like + Reddit-like     (ReportAgent with
                                   social networks)       parallel interaction)            deep analysis tools)

Phase	What happens
Graph Build	Extracts entities, relationships, and context from your documents. Builds a knowledge graph via GraphRAG.
Environment Setup	Generates agent personas with distinct personalities, beliefs, and social connections.
Simulation	Agents interact across dual platforms (Twitter-like and Reddit-like) in parallel. Dynamic memory updates each round.
Report Generation	A ReportAgent analyzes the post-simulation environment — sentiment shifts, faction formation, viral dynamics, outcome trajectories.
Deep Interaction	Chat with any agent to understand their reasoning. Query the ReportAgent for follow-up analysis.

Quick Start

1. Get an API key

Sign up at deepmiro.org → Dashboard → API Keys. Your key looks like dm_xxxxxxxxx.

2. Install

Pick the install path for your client. Don't install the .mcpb desktop extension if you're using Claude Code or Claude Cowork — those need the plugin to get the /predict skill, background polling, and live narration.

Claude Desktop → use .mcpb

1. Download deepmiro.mcpb from the latest release
2. Claude Desktop → Settings → Extensions → Advanced settings → Install Extension → pick the file
3. Paste your API key when prompted

Claude Code & Claude Cowork → use the plugin

The plugin ships the /predict skill — the MCP alone is missing the orchestration logic (background polling via cron, live agent narration, the setup wizard).

claude plugin marketplace add kakarot-dev/deepmiro
claude plugin install deepmiro@deepmiro-marketplace
export DEEPMIRO_API_KEY=dm_your_key   # or set in ~/.claude/settings.json

Restart Claude Code, then say /predict or predict how people will react to [scenario].

Everywhere else → npm package

Generic MCP install for clients that aren't Claude Desktop, Claude Code, or Claude Cowork:

Client	Install
OpenAI Codex	codex plugin install kakarot-dev/deepmiro
ChatGPT Desktop	Settings → MCP Servers → Add → npx deepmiro-mcp with env DEEPMIRO_API_KEY
Cursor / Windsurf	Settings → MCP → Add → npx deepmiro-mcp with env DEEPMIRO_API_KEY
VS Code (Copilot)	Add to .vscode/mcp.json: "deepmiro": {"command": "npx", "args": ["-y", "deepmiro-mcp"], "env": {"DEEPMIRO_API_KEY": "dm_xxx"}}

Rehearse the Future in 60 Seconds

Four services, one compose file, one API key.

What gets deployed

Service	Role
backend	Flask engine that runs the OASIS multi-agent simulations
mcp	Public entry point for AI tools (Claude, Cursor, VS Code)
twhin-sidecar	Shared TWHIN-BERT embedding service (loads once per pod)
surrealdb	Graph + vector + document store for agents and reports

What you need

• A Fireworks AI key (fireworks.ai, ~$5 free credit) — covers primary LLM, boost, and embeddings in one key. Any OpenAI-compatible API also works.
• openssl rand -hex 32 for your SurrealDB root password.
• ~$5–10/month of Railway credit if you're using the template.

Option A — Railway one-click

Railway reads docker-compose.yml from the repo root and prompts for LLM_API_KEY + SURREAL_PASSWORD. The MCP service gets a public *.up.railway.app URL — hand that to your AI tools.

Note — LLM provider on Railway. The template ships with Fireworks as the default (primary: minimax-m2p5, boost: gpt-oss-120b, embeddings: nomic-embed-text-v1.5 — one key covers all three). Any OpenAI-compatible API works — to swap to OpenAI, Together, Groq, Ollama, vLLM, or anything else, change LLM_BASE_URL and LLM_MODEL_NAME on the backend service's Variables tab after deploy. Same for LLM_BOOST_* if you want a separate reasoning model, and EMBEDDING_* if you want a different embedding provider.

Option B — Docker (self-hosted)

git clone https://github.com/kakarot-dev/deepmiro.git
cd deepmiro && cp .env.example .env

Edit .env — two required variables:

LLM_API_KEY=your-fireworks-or-openai-key
SURREAL_PASS=$(openssl rand -hex 32)

Start everything:

docker compose up -d

This pulls pre-built images from GHCR and starts four services:

Service	Port	Description
mcp	3001 (public)	MCP server — the only exposed port
backend	5001 (internal)	Flask simulation engine
surrealdb	8000 (internal)	Graph + vector store
twhin-sidecar	7001 (internal)	Shared TWHIN-BERT embeddings

First startup takes ~2 minutes (TWHIN-BERT model warm-up). Check readiness:

docker compose logs -f twhin-sidecar   # wait for "TWHIN-BERT ready"
docker compose logs backend            # wait for "MiroFish Backend 启动完成"

To build from source instead of pulling images:

docker compose -f docker-compose.yml \
  --build \
  -f docker/Dockerfile.backend \
  up -d

Or uncomment the build: blocks in docker-compose.yml and comment out the image: lines.

MCP lives on http://localhost:3001. Backend and SurrealDB stay internal to the compose network unless you explicitly publish them (see docker-compose.yml comments for how).

Wire it into Claude Desktop

Add DeepMiro to your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

For a local Docker deployment:

{
  "mcpServers": {
    "deepmiro": { "url": "http://localhost:3001/mcp" }
  }
}

For Railway or any public deployment:

{
  "mcpServers": {
    "deepmiro": { "url": "https://your-app.up.railway.app/mcp" }
  }
}

Restart Claude Desktop after editing. Then ask: "Use DeepMiro to simulate how 100 senior engineers would react to a return-to-office mandate" — and paste the memo.

What it costs

• Railway: ~$5–10/month (four services, ~4 GB resident)
• LLM: ~$0.10–0.20 per quick-preset simulation on Fireworks
• TWHIN-BERT: zero — runs locally in the sidecar

Security

MCP ships with no auth by default — set MCP_API_KEY in .env before exposing it to the internet. The backend REST API is internal-only out of the box.

Skip the deploy entirely? Use the hosted version at deepmiro.org — same engine, same models, no Docker.

MCP Server

DeepMiro is an MCP server. MCP is the universal standard adopted by Claude, ChatGPT, Gemini, Cursor, VS Code, and every major AI client — one server, works everywhere.

npx deepmiro-mcp

Available tools: create_simulation, simulation_status, get_report, interview_agent, upload_document, list_simulations, search_simulations, simulation_data, cancel_simulation.

What's Different

DeepMiro is a performance-focused fork of the original MiroFish engine. Same OASIS simulation core, rebuilt infrastructure:

Component	MiroFish (original)	DeepMiro
Recommendation engine	Full LLM call every round (~200s/round)	Cached TWHIN-BERT embeddings (~15ms/round)
Entity extraction	Sequential NER	5-worker parallel NER via ThreadPoolExecutor
Graph build time	~5 minutes	~56 seconds
Graph database	Zep Cloud (proprietary)	SurrealDB (self-hosted, open-source)
Vector search	Cloud-dependent	Hybrid HNSW + BM25 (local, 768-dim cosine)
Embedding model	Tied to Zep	nomic-embed-text-v1.5 via Fireworks (swappable)
Document ingestion	Manual text input	Upload endpoint with magic-byte validation (PDF, MD, TXT)
LLM provider	Alibaba Qwen (hardcoded)	Any OpenAI-compatible API
Deployment	Docker only	Docker + Helm chart + k3s-ready

Persona Fidelity: How DeepMiro Keeps Agents In Character

Multi-agent LLM simulations have a dirty secret: personas drift. By round 20, Tucker Carlson starts quoting the ACLU. By round 45, Marco Rubio sounds like Bernie Sanders. Every distinct voice collapses into the same bland "helpful assistant" register.

This isn't a prompting problem — it's an attention decay problem. Kim et al. (COLM 2024) proved that LLM attention to system-prompt tokens decays geometrically over turns. LLaMA2-70B drifts significantly within 8 turns. Larger models drift more, not less. A 2KB persona cannot compete with 50KB of accumulated conversation history.

Every naive multi-agent simulation hits this wall. DeepMiro doesn't, because we copied what Stanford's Generative Agents (Park et al. 2023) did for their 25-agent Smallville simulation — with some practical shortcuts.

What we do

1. Structured personas with explicit negative examples. Every agent gets a structured profile alongside the prose bio:

• ideology_anchor — a 2-5 word partisan tag ("conservative populist", "progressive labor")
• core_beliefs — 3-5 first-person declarative statements, no hedging
• verbal_tics — 3-5 literal phrases the person actually uses
• never_say — 3-5 sentences the person would refuse to utter
• speaking_style — register + rhetorical habits

The never_say block is the drift killer. Models drift toward the centroid of what they say. Explicit negative examples ("Tucker Carlson would never say 'I stand with the ACLU'") anchor the LLM against that collapse.

2. Dynamic persona regeneration per round. Instead of locking the persona in at the system-prompt level and watching attention decay from round 1, we rebuild system_message.content before every agent acts. Each round, the agent sees a fresh third-person character brief:

# Character Brief: Tucker Carlson

The agent in this conversation is Tucker Carlson.
You are simulating how Tucker Carlson would respond.

## What Tucker Carlson Would NEVER Say
- "I stand with the ACLU"
- "We need to find common ground with progressives"
...

## What Tucker Carlson Has Said Recently
- "Permanent Washington wants you to believe..."
- "Let's pause for a moment — they're not even hiding it"
...

## Task
What would Tucker Carlson actually do? React in his authentic voice.
Do not become a neutral assistant. Do not seek balance.

The persona never gets stale because it's built fresh from the same structured fields every turn.

3. Third-person framing. "You are Tucker Carlson" triggers RLHF helpful-assistant sycophancy — the model tries to be polite and balanced because that's how it was trained to respond to "you are X" instructions. Third-person framing ("the agent is Tucker Carlson", "what would Tucker Carlson do?") bypasses that trigger entirely. This single change is load-bearing.

4. Self-consistency anchor. Each round injects the agent's own 3 most recent posts as reference material. Tucker Carlson sees what he just said, which makes him more likely to say something consistent with it. This is cheap drift resistance — no extra LLM calls, just reading from the action log.

5. No accumulated chat history. Unlike naive multi-agent setups, DeepMiro does NOT feed each agent the rolling conversation history from previous rounds. Agents get their fresh persona + the current feed observations. Attention stays focused on character + present context, not on 50KB of stale noise.

What we don't do

• We don't script reactions. Agents aren't told "mock liberal content" or "support conservative content" — that would script the outcome and destroy the simulation's predictive value. The emergent behavior is the whole point.
• We don't filter feeds by ideology. Tucker Carlson sees AOC's posts. That's how he has something to push back against. Echo chambers are not simulations.
• We don't fork OASIS. The entire fix is a runtime wrapper around CAMEL's agent pager. No upstream drift, no fork maintenance.

Research foundations

Technique	Source
Attention decay over system prompts	Kim et al. — Measuring and Controlling Persona Drift (COLM 2024)
Third-person framing bypasses RLHF sycophancy	Park et al. — Generative Agents (Stanford 2023)
Negative examples > positive instruction	Examining Identity Drift in LLM Agents (arXiv 2412.00804)
Dynamic persona summary per action	Park et al. — Generative Agents (Stanford 2023)
JSON personas collapse to neutral register	Persona-Aware Contrastive Learning (ACL 2025)

Benchmarks

15-agent quick simulation, enriched prompt, measured end-to-end:

Stage	Time
Graph build	~10s
Agent generation	~3 min
Simulation (110 Twitter + 26 Reddit actions)	~4 min
Total pipeline	~7 min (quick) / ~12 min (standard, 80 agents)

The biggest win is the recommendation system: TWHIN-BERT embeddings are computed once per user at setup, then only new posts are embedded incrementally each round. Cosine similarity via numpy replaces what was previously a full LLM inference call — 13,000x faster per round.

Monorepo Structure

deepmiro/
├── engine/              # Python Flask simulation backend
│   ├── app/
│   │   ├── api/         # REST endpoints (simulation, graph, documents, report)
│   │   ├── services/    # Graph builder, simulation runner, report agent
│   │   ├── storage/     # SurrealDB adapter, embedding service, NER
│   │   └── utils/       # LLM client, retry logic, logging
│   └── pyproject.toml
├── mcp-server/          # TypeScript MCP server (npm: deepmiro-mcp)
│   └── src/
├── .claude-plugin/      # Claude Code plugin + marketplace manifests
├── .codex-plugin/       # OpenAI Codex plugin manifest
├── .agents/             # Codex marketplace catalog
├── .mcp.json            # MCP config (auto-loaded when running `claude` here)
├── skills/predict/      # /predict skill (auto-setup, narration, interviews)
├── helm-chart/          # Kubernetes (k3s) deployment
├── docker/              # Dockerfiles + compose
├── docs/                # Landing page
└── locales/             # i18n (en, zh)

Use Cases

Domain	Example
Market analysis	Upload an earnings report. "How will retail investors react to this guidance revision?"
Policy testing	Upload a draft regulation. "What public backlash should we expect, and from which demographics?"
PR & comms	Upload a press release. "How will this announcement play on social media over 48 hours?"
Competitive analysis	Upload competitor product specs. "How will our user base respond to this feature gap?"
Creative exploration	Upload a novel's first 80 chapters. "What ending would emerge from these character dynamics?"
Crisis simulation	Upload an incident report. "How does public opinion evolve if we respond with X vs Y?"

Acknowledgments

DeepMiro is a fork of MiroFish, originally created by Guo Hangjiang and supported by Shanda Group. The simulation layer is powered by OASIS from the CAMEL-AI team.

License

AGPL-3.0

---

deepmiro.org · Built by Joel Libni