What is Yggdrasil Agent?

Yggdrasil Agent is a vendor and model agnostic advanced reasoning system that explores multiple reasoning paths in parallel across ten cognitive realms, scores intermediate steps, resolves conflicts, and converges on one validated answer. Access via REST API, MCP gateway, CLI, or Chainlit web UI.

How is Yggdrasil different from LangGraph, CrewAI, or AutoGen?

Unlike LangGraph, CrewAI, or AutoGen, Yggdrasil ships parallel multi-agent reasoning out of the box: ten realms plus Four Stags (multi-agent debate) per realm, Valkyrie early exit, ChromaDB-backed memory layers, and 76+ MCP tools via Brokkr without custom graph or crew design.

What models and deployment options does Yggdrasil support?

Yggdrasil supports Claude, Gemini, GPT, DeepSeek, Qwen, Mistral, and GLM, routed by role via OpenRouter, direct provider APIs, or Ollama for local deployment. Docker Compose runs cloud frontier, open-weights, or fully local modes with Chainlit UI, CLI, and FastAPI REST API.

What is Yggdrasil's GAIA benchmark score?

Yggdrasil Agent achieves 71.7% overall accuracy on the GAIA benchmark (43/60 questions) by utilizing parallel reasoning paths. Level 1: 85%, Level 2: 85%, Level 3: 45%, self-measured on Yggdrasil's own harness with 20 questions per level.

Can Yggdrasil run fully locally with Ollama?

Yes. Yggdrasil supports a full local deployment mode with Ollama on your hardware so data stays on your network. The same agent stack, memory profiles, and interfaces (CLI, Chainlit UI, REST API) run in cloud, open-weights, or local configurations.

Yggdrasil Agent: Multi-Agent AI Reasoning Framework

See It In Action

The Chainlit web UI and CLI share one workspace via Meta-Bifröst (unified workspace bridge). Web: pick a Realm, ask anything, expand Yggdrasil Reasoning. CLI: parallel realms, live streaming, cited final answers with cost and source stats.

CLI demo (Odin mode): using parallel realms, web search, streaming thoughts, and tool calls as the answer builds.

CLI full answer: full output from the Odin mode demo above, fibromyalgia research with citations, run stats, and sources.

Yggdrasil Agent web UI home screen with Ten Realms picker and Web, Tools, and Images toggles — Web UI home: parallel reasoning intro, Ten Realms picker, and Web, Tools, and Images toggles.

CLI demo (no parallel): single-realm run for comparison, same CLI, with parallel execution disabled.

Chat UI demo: Fast Path in seconds. Expand the trace for Answer Extraction, realm routing, and the full chain.

Real-World Use Cases

Real-world use cases for Yggdrasil Agent include software development, research, data science, business analysis, and multilingual work. Yggdrasil Agent adapts to your task: from spreadsheet analysis to citation-heavy research.

Software Dev

Debug production issues, code review, architecture design, security audits

Research

Literature reviews, multi-language sources, citation-heavy reports

Data Science

Excel/CSV analysis, statistical modelling, trend prediction with Sindri (autonomous coding agent)

High Quality Research Sources

Evidence-based sources, PubMed priority, peer-reviewed evidence valued, trusted guidelines and citations

Business

Market analysis, competitive intelligence, strategic planning

Multilingual

Translation, cross-language research, localisation (12+ languages)

Weather / Aura

Conditions, forecasts, and ambient context for planning and mood

Brainstorming

Ideation, alternatives, creative exploration with parallel branches

Learning

Explain concepts, summaries, Q&A, and step-by-step tutorials

Writing / Content

Drafts, outlines, editing, and long-form content with cited sources

Math & STEM

Proofs, physics, formal logic. Kvasir (verified math reasoning mode) with DeepSeek-R1 and ToRA verification

Document / OCR

Invoice and receipt processing, form extraction, scanned PDFs with Veðrfölnir vision

Beyond the Chatbot

A typical AI answers once and moves on. Context bloats. Wrong paths finish. Yggdrasil explores from multiple angles, checks its own reasoning, and lands on one answer it can account for.

Where Yggdrasil Fits

Some tools code. Some deliver work. Some automate apps.
Yggdrasil consolidates those strengths in one stack, then validates the answer.

Different jobs. One stack.

Research breadth Manus · Deep Research Broad research agents. Sources, subtasks, one merged report.

Finished deliverables Perplexity Computer Cloud digital worker. Describe the deliverable. Models, connectors, background runs.

Coding loops Cursor · Codex · Claude Code Repo coding loops. Edits, refactors, terminal work, pull requests.

Personal automation Hermes · OpenClaw Self-hosted personal agents. Messaging, apps, skills that compound.

One Stack. Validated Answer.

Research, tools, code, files, connectors, long sessions. Pulled into one place.

Research: sources. Workers: routing. Code: sandboxes. Personal: memory.

◆ Multiple perspectives: evidence, synthesis, challenge, and first principles in parallel

◆ First-principles checks: what is known, what breaks, what would change the answer

◆ Conflict scoring: evidence strength, source independence, real separation between answers

◆ Echo-chamber checks: same sources do not count as independent agreement

◆ Cross-provider judging: scoring spans multiple model providers

◆ Self-fixing retries: weak drafts retried and checked against evidence

◆ Failure memory: remembers dead ends, not just preferences

◆ Fuller field view: evidence, conflicts, tradeoffs, and why this answer

GAIA scores vary widely by harness. The same model can swing by tens of points on tooling alone (see Princeton HAL). Public leaderboards mix bare models, custom scaffolds, and managed agents. Any figures here are self-measured on Yggdrasil's own harness and offered as a reference point, not a ranking. Compare architectures and reproducibility, not headline numbers.

Choose Your Mode

Switch modes via the icons in the chat or type slash commands: /loki, /odin, /kvasir, /tyr, /sindri, /export.

Normal

Balanced reasoning (default). ~7 pages, 100-300s. Resets persona modes. Most everyday tasks and questions.

Loki

Loki (fast-chat persona mode): ultra-fast chat with witty persona. Bypasses deep reasoning for instant responses (~1-3 sentences, <1s). Quick answers, brainstorming, casual chat.

Kvasir

Kvasir (verified math reasoning mode): SOTA math reasoning with ToRA verification. Proofs, physics, formal logic. ~15 pages, 150-500s.

Tyr

Tyr (multilingual research mode): research with sources in 12+ languages. Priority on primary sources. ~10 pages. International perspectives.

Eir

Eir (science and health research mode): academic citations and interrogative workflow. PubMed priority. Peer-reviewed evidence valued. ~10-15 pages.

Odin

Odin (deep multi-realm analysis mode): multi-realm deep analysis with Battle Plan display and subtask progress. ~20 pages, 400-800s. Complex problems, strategic planning.

Sindri

Sindri (autonomous coding agent): autonomous coding with Claude Code in E2B sandbox. Multi-turn dev work. Variable length (60s-20min). Coding, refactoring, file analysis.

Built For How You Work

Yggdrasil Agent is a vendor and model agnostic reasoning system. Here is what matters to each kind of user, and how the system is designed to help: no lock-in, broad customization, and tools to manage your own data.

Researchers & Analysts

Need: answers they can trust and cite.

Tyr (multilingual research mode): sources in 12+ languages
Eir (science and health research mode): peer-reviewed, PubMed-priority evidence
Fimbulwinter (multi-source verification pipeline): multi-source verification
Paper-style PDF export with clickable citations

Data Scientists

Need: run real code on real data.

Sindri (autonomous coding agent): Python, shell & spreadsheets in an E2B sandbox
Kvasir (verified math reasoning mode): math, proofs, and formal logic with verified steps
Veðrfölnir (centralized vision / OCR): OCR & tables from PDFs and images

Decision-makers & Operators

Need: predictable cost and speed.

Valkyrie (breakthrough early-exit optimizer) + parallel scouts (up to 2.2× faster)
Real-time USD cost tracking
Smart context truncation cuts tokens, cost, and latency

Teams Who Want Flexibility

Need: no vendor or model dictating the stack.

Vendor agnostic: OpenRouter, direct provider APIs, or Ollama locally
Model agnostic: Claude, Gemini, GPT, DeepSeek, Qwen, Mistral, GLM, and more, routed by role, not one default model
Modes, personas & memory profiles you can tune

Three Deployment Modes

Need: cloud power, open models, or full local control.

Cloud frontier: full cloud with SOTA frontier models via direct provider APIs
Open weights: open-weight models routed through OpenRouter with the same agent stack
Local with Ollama: alternative full local mode on your hardware, data stays on your network

Service Backends

Need: swappable backends per role, not one bundled vendor.

Web search: Perplexity Sonar (not Perplexity Computer)
Code execution: E2B sandbox, Claude Code via Sindri (autonomous coding agent)
Authentication: Google OAuth, JWT sessions
Assignments in yggdrasil.toml, updatable as services ship

Features

Every feature targets a real failure mode of single-LLM tools.

The Four Stags (Multi-Agent Debate Framework) Learn more →

The Four Stags (multi-agent debate framework) kills single-perspective blind spots: four agent perspectives challenge every realm in parallel, so weak ideas get caught instead of confidently shipped.

❇

Dáinn, The Chronicler (Evidence Gatherer)

Gathers evidence and prior knowledge. Best for research and building on established work.

✴

Dvalinn, The Weaver (Synthesis Agent)

Threads ideas into unified narratives. Best for synthesis and finding coherence across sources.

☸

Duneyrr, The Challenger (Adversarial Critic)

Questions assumptions and probes contradictions. Best for critical analysis and stress-testing.

✵

Durathor, The Pathfinder (Creative Explorer)

Explores unexpected solution spaces. Best for creative breakthroughs and novel angles.

The World Tree

The World Tree (parallel multi-agent architecture) is Yggdrasil Agent's core design. Ten realms (parallel cognitive modes) reason in parallel while specialized agents route the work, score every thought, resolve conflicts, validate output, and learn across sessions, so the system stays fast, accurate, and interpretable. Four Stags (multi-agent debate) →

Ten cognitive modes. Hover or tap a realm to see how it thinks.

The ten realms map to validated cognitive modes from Tree-of-Thoughts, Graph-of-Thoughts, and cognitive science. Read the architecture article →

𐂃

Sleipnir

Sleipnir (realm router): routes to next realm or Ragnarök every iteration

👁

Odin's Foresight

Odin's Foresight (task decomposer): task decomposition, GAIA patterns, battle plans

⚒

Brokkr

Brokkr (MCP tool selector): smart tool selection across 76+ MCP tools

𓅇

Veðrfölnir

Veðrfölnir (centralized vision): vision extraction — OCR, layout, one pass for all realms

ᚾ

Sindri

Sindri (code execution agent): code execution in E2B sandbox; Python, shell, spreadsheets

𔒝

Muninn

Muninn (long-term memory / RAG store): recall and store across sessions. Memory Vault: stats, backup, ChromaDB download.

𒀀

Norns

Norns (thought scorer): score every thought for relevance, evidence, and goal fit to rank and prune paths

⚖

Forseti

Forseti (conflict resolver): conflict resolution with 3-pillar LLM-as-Judge

𖣑

Heimdall

Heimdall (insight bridge): insight bridge between parallel agents

⛓

Gleipnir

Gleipnir (loop prevention): loop prevention, circuit breakers, forced convergence

𓆩✧𓆪

Valkyrie

Valkyrie (breakthrough early-exit optimizer): breakthrough-triggered early exit on parallel scouts

Ragnarök

Ragnarök (final synthesizer): final synthesis from the best thoughts across all realms

You ask Recalls what's relevant Focused working memory Answers Saves useful learnings

Memory that helps, not hoards

Profiles · Private mode · Fresh start

Efficiency · Cross-session recall · Learning · Your data, your control

❟❛❟

Save & rewind

Save your progress. If something goes wrong, rewind and try again, no starting from scratch.

◆ Saves at important steps

◆ Rewinds when something fails or times out

◆ Tries again with a different approach

◆ No redoing everything from the start

𔓂

In-run evidence (Vitni session cache)

During one answer: one search, every realm benefits. No repeated lookups.

◆ Findings are shared across all reasoning paths

◆ No duplicate web searches or tool runs

◆ Smart recall: similar questions reuse answers

◆ Fresh each session: resets after 24 hours

ᚺ

Learn from mistakes

Helheim (failure memory store) remembers what went wrong. Doesn't repeat the same mistakes.

◆ Remembers why something failed

◆ Across conversations: lessons stick around

◆ Learns from every dead end

Manage your data

Export and download your data, and clear your history when you want.

◆ Download all data as files

◆ Clear history: wipe sessions and memory on demand

◆ Privacy modes: normal, private, no memory, or fresh

◆ Trace masking on by default

Usage & Integrations

Usage options for Yggdrasil Agent are three: interactive web UI, command-line with memory profiles, or headless REST API. All three share the same core reasoning engine. Meta-Bifröst (unified workspace bridge) unifies CLI, web UI, and REST API in one workspace.

Chainlit Web UI

The primary interface. Chat with streaming, visual realm navigation, creature insights, real-time cost tracking (USD), voice input, and PDF export of reasoning sessions with clickable citations.

# Docker service
docker compose up yggdrasilagent-ui
# Then open:
http://localhost:8000

◆ Streaming chat with realm progress

◆ Google OAuth (production) or local auth

◆ Typo-tolerant quick commands

◆ Real-time cost summary per session

◆ PDF export (scientific paper-style, clickable citations)

CLI Interface

Full-featured command-line agent with memory profiles, session management, persona switching, privacy modes, and budget-constrained queries.

# Web search + memory profile
docker compose run --rm yggdrasilagent-cli \
yggdrasil --web-search --profile work \
"Analyze AI agent market 2026"
# Odin persona, 8 iterations
docker compose run --rm yggdrasilagent-cli \
yggdrasil --persona odin --max-iterations 8 \
"Your question"

◆ Privacy: --private recalls past sessions but doesn't save this one; --fresh is a one-off with no memory (all modes)

◆ Launch multiple agents in the background

◆ Session resume & export

REST API (FastAPI)

Headless access via Google OAuth2 + JWT. Full Swagger UI at /docs. Built for your apps and workflows.

POST /api/v1/reason
{
"query": "Analyze EV market 2026",
"max_iterations": 5,
"enable_web_search": true,
"profile": "business"
}
# → final_answer, cost, citations

◆ Google OAuth2 + JWT auth

◆ Session management endpoints

◆ Swagger UI at /docs

Security & Privacy

Don't fear autonomous agents. Govern them. Yggdrasil Agent includes human-in-the-loop gates, output validation, trace masking on by default, and tools to manage your own data, with private deployment options for enterprise.

Memory & Privacy Modes

Pick how much Yggdrasil remembers and stores. Use separate profiles (--profile work) to keep work and personal contexts apart.

Mode	Remembers past sessions	Saves this session
Normal	Yes	Yes
Private	Yes	No
No memory	No	No
Fresh	No (temporary)	No

CLI: --private, --no-memory, --fresh · Web UI: privacy toggles in chat settings

🛡

Action Gates (HITL)

Sensitive tool calls (filesystem, shell, external transfers) pause for mandatory user approval before they run.

🧾

Trace Masking

Secret and personal-data masking in observability traces is on by default, helping keep sensitive data out of logs.

Auth & Isolation

Google OAuth2 and JWT-based session isolation keep each user's sessions and memory separate.

⚖

Guardrails

Baldr output validation, Mimir claim verification, prompt-injection detection, and a first-person crisis guardrail.

Data Controls

Export and download conversations, memory, and sessions. Privacy modes let you query without persisting.

☁

Private Deployment

Fully dockerized. Hosted by default, with dedicated and private deployment options, including your own cloud.

Research Foundation

Yggdrasil Agent's parallel reasoning builds on peer-reviewed work in multi-path inference, self-improvement, and long-term memory, measured continuously against the GAIA benchmark as a fixed point of reference.

Tree of Thoughts

Yao et al., 2023: multi-path exploration across realms and stags.

Graph of Thoughts

Besta et al., 2023: non-linear reasoning and thought aggregation.

Reflexion

Shinn et al., 2023: self-improvement via reflection (Huginn reflection agent).

Generative Agents / MemGPT

Park et al. & Packer et al., 2023: memory importance and consolidation (Muninn long-term memory store).

GAIA Benchmark Methodology

GAIA guide →

Measured on the GAIA benchmark, 20 questions per level. Accuracy is taken from GAIA result JSONs; cost combines LLM spend (from Langfuse traces) with web-search cost.

Level 1

85%

17/20 correct
~2.0 min · ~$0.07

Level 2

85%

17/20 correct
~5.3 min · ~$0.07

Level 3

45%

9/20 correct
~6.2 min · ~$0.09

Overall

71.7%

43/60 questions

See It In Action

Real-World Use Cases

Open source, this winter

Beyond the Chatbot

Where Yggdrasil Fits

Choose Your Mode

Built For How You Work

Researchers & Analysts

Data Scientists

Decision-makers & Operators

Teams Who Want Flexibility

Three Deployment Modes

Service Backends

Features

The Four Stags (Multi-Agent Debate Framework) Learn more →

The World Tree

Save & rewind

In-run evidence (Vitni session cache)

Learn from mistakes

Manage your data

Usage & Integrations

Chainlit Web UI

CLI Interface

REST API (FastAPI)

Choose Your Tier

Security & Privacy

Memory & Privacy Modes

Research Foundation

GAIA Benchmark Methodology

Plant the Seed Today

Custom mode & consultation

See It In Action

Real-World Use Cases

Open source, this winter

Beyond the Chatbot

Where Yggdrasil Fits

Choose Your Mode

Built For How You Work

Researchers & Analysts

Data Scientists

Decision-makers & Operators

Teams Who Want Flexibility

Three Deployment Modes

Service Backends

Features

The Four Stags (Multi-Agent Debate Framework) Learn more →

The World Tree

Save & rewind

In-run evidence (Vitni session cache)

Learn from mistakes

Manage your data

Usage & Integrations

Chainlit Web UI

CLI Interface

REST API (FastAPI)

Choose Your Tier

Security & Privacy

Memory & Privacy Modes

Research Foundation

GAIA Benchmark Methodology

Parallel AI Reasoning Architecture

Plant the Seed Today

Custom mode & consultation