Name: Yggdrasil Agent
Availability: PreOrder
Author: Bartosz Lenart

Advanced Reasoning System A private AI that thinks in parallel, learns, adapts, and stays with you everywhere you work

Depth: It takes multiple cognitive perspectives and shows the evidence behind the winner, grounded in high-quality sources.
Integrity: One AI across editor, terminal, browser, and apps, with shared memory that grows with you.
Freedom: Works with infinite models and tools: paid cloud or free local, any MCP tool or CLI. Pick per task for quality, cost, or privacy.
Ownership: Own your data and architecture: download or delete anytime, run fully local when privacy matters, and extend it programmatically.

Preview this summer 2026 with $5 in free credits.

I'm also looking for early contributors.

Core open source in winter.

Independent project by

Bartosz Lenart

Watch The Work,
Not A Mockup

Real recordings of the product answering hard questions. Watch it think, cite, and price every run.

Yggdrasil Agent web UI home with Tree of Thoughts tagline and Web, Tools, and Images toggles — Web UI home: Tree of Thoughts landing with toggles.

Yggdrasil Agent web UI Ten Realms picker with Alfheim selected and suggested prompts — Web UI home: Tree of Thoughts landing with toggles.

Chrome extension: side chat on any page: generated images, File Vault, and page-aware and execution capabilities answers in place.

Chat UI demo (self-hosted cloud via web): Fast Path in seconds. Expand the trace for Answer Extraction, realm routing, and the full chain. Image and video generation, operation on files accessed via mobile app.

PDF report (normal mode): sample research output with clickable citations.

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 CLI demo above, fibromyalgia research with citations, run stats, and sources.

Connect Yggdrasil to your coding tool: sidebar extension in VS Code and Cursor with all reasoning modes, file handling, @context attach, access to memories and file vault.

One system, one memory,
Multiple interfaces

Infinite adaptation capabilities. Every window into the same core.

CLI Interface

Full-featured command-line agent on the same Docker server as Bifrost for VS Code: memory profiles, sessions, personas, 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 / --fresh modes
Background multi-agent runs
Session resume & export

Chainlit Web UI

Chat with streaming, visual realm navigation, creature insights, real-time cost tracking (USD), voice input, and PDF export with clickable citations. Runs locally in Docker today.

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

Streaming chat with realm progress
Google OAuth (production) or local auth
Real-time cost summary per session
Cloud at yggdrasilagent.com/ai via the waitlist (soon)

REST API & MCP

Call the same reasoning engine from your apps. Local Docker today; cloud on the waitlist.

REST API POST /api/v1/reason with OAuth2 + JWT. Docs at /docs.

MCP Install into Claude, Cursor, and other coding agents with yggdrasil mcp install. They call the same realms as one tool.

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

Bifrost for VS Code & Cursor

Sidebar agent that reads your working directory and talks to the same Docker server as the CLI. Hand hard work off to the cloud when you need more power.

Workspace aware Understands the content of your working directory so answers stay grounded in the project you have open.

Same Docker server Same local Docker backend as the CLI (and every other interface), or cloud for heavier tasks.

Works in VS Code and Cursor
Project-aware chat from the sidebar

Bifrost for Chrome

A side panel on any page that reads the tab you are on: page-aware research, shopping comparisons, and product Q&A without leaving the browser.

Page context Reads the active tab so answers stay grounded in what you are viewing.

In-page control Ask about products, compare options, and research without copy-pasting into another app.

Chrome side panel on any site
Page-aware research and shopping Q&A

Open Design & ACP editors

Design and build artifacts inside a real design app. Yggdrasil registers as the agent runtime over the Agent Client Protocol, so reasoning, tool approval, and the files it writes stay in the project you have open.

# Open Design with Yggdrasil as the agent
docker compose -f docker-compose.yml \
-f docker-compose.opendesign.yml up open-design
# Then open:
http://localhost:7457

Live preview & export Every write reports its real path, so the preview pane updates as the artifact changes and HTML, PDF, or PPTX export matches what you see.

Your project, not scratch space Only files inside the folder you opened count as project files; Yggdrasil's internal storage stays out of the design.

Zed and marimo speak the same yggdrasil acp surface
Permission prompt in the host before a risky tool runs
Same thread and memory as CLI, web UI, and VS Code

Scheduler

Unattended runs from plain-English schedules. The same realms fire while you are away, and the answer lands in a normal conversation you can open later.

# Interval
every 30m · every 2 hours · every day
# Wall-clock
daily at 9am · weekdays at 18:30 · every monday at 9:00
# One-shot
2026-08-01 09:00 · tomorrow at 14:00 · in 2 hours · at 15:30
# Advanced (optional)
cron: 0 9 * * 1-5

Interval & wall-clock every day is 86400s. Daily and weekday times stay on the clock (no drift).

One-shot & cron at 15:30 means today if still ahead, else tomorrow. Supported cron expands to daily + days; unsupported forms get a clear error.

Opt in with [scheduler] in yggdrasil.toml
Results keep their thread across CLI, web UI, and other interfaces

Messaging Soon

Chat apps and bots on the same reasoning backend: Slack, Teams, Telegram, and more, without leaving the thread you already use.

Chat apps & bots Bring Yggdrasil into the channels where your team already works.

Same engine Realms, memory, and Valkyrie early exit, from a DM or a group channel.

Coming soon: messaging surfaces for chat apps and bots
Same modes and reasoning backend as web, CLI, and IDE
Designed for self-hosted personal agents

Start with Docker Compose, pick an interface, keep the same reasoning backend and the same conversation: a thread you start in the terminal continues in the browser. Every service uses the Docker MCP Toolkit: the most convenient way to handle and manage tools.

One Engine
to Rule Them All

Great tools already exist for research, coding, and automation.
Yggdrasil brings those strengths into one system and adds the part they miss: a validated answer.

What each tool is best at

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.

Yggdrasil: One System, A Validated Answer

Lives in the web UI, REST API, and VS Code. Understands web pages and code, runs deep research, and keeps files, tools, and long sessions in one place. Built to verify claims against relevant sources, with scientific sources prioritized when they matter.

A typical chatbot follows one line of thought, sounds sure when sources echo, and hides the steps. Yggdrasil overlaps with all four above, then does the opposite and adds what none of them have:

◆ Think in parallel: many independent paths at once, not one early guess that hardens into certainty

◆ Trust the evidence: claims face real sources; science wins when it should

◆ Make them fight: critics and conflict checks kill weak answers before a winner is chosen

◆ Fix before you see it: bad drafts die in the lab, not in your chat

◆ Remember the scars: failed paths stay marked so they are not walked again

◆ Show your work: citations, cost, and the full trail for why the winner won

Where Multiple Perspectives Matter

Flagship outcomes first. Use Yggdrasil when one plausible answer is not enough, then dig into architecture and usage for the full stack.

Software Dev

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

Research

Literature reviews across 12+ languages, multi-source claim checks, paper-style PDF export with clickable citations

Data Science

Excel/CSV analysis, stats, and trend work with real code in a secure sandbox (Sindri)

Business

Market and competitive analysis with live USD cost tracking and early exit when the answer is clear

Writing / Content

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

Math & STEM

Proofs, physics, and formal logic with verified steps (Kvasir)

Document / OCR

Invoices, receipts, forms, and scanned PDFs: OCR and tables in one pass (Veðrfölnir)

Browsing & Shopping

Bifrost Chrome extension: a side-chat window on any page. Reads and understands the page you're on for shopping comparisons, product research, and page Q&A

Next: how you steer it, how it thinks, then the deep dive: Modes, Architecture, and the long-form article.

Choose Your Mode

Same persona icons as the live chat. Pick a specialist for speed, research, science, math, deep analysis, or coding, or leave Normal for balanced everyday work.

Normal

Balanced reasoning for everyday work, typically 100–300 seconds. Select a specialist mode for speed, research, science, math, deep analysis, or coding.

How The System Works

The question travels the World Tree: split, explored on several paths, scored, and merged into one answer you can inspect. Inside each path, roles gather evidence, merge ideas, attack weak claims, and explore alternatives.

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

Sleipnir

realm router

Routes to the next realm or Ragnarök every iteration.

Odin's Foresight

task decomposer

Splits work into GAIA patterns and battle plans.

Brokkr

MCP tool selector

Smart tool selection across 76+ MCP tools.

Veðrfölnir

centralized vision

OCR, layout, and vision in one pass for every realm.

Sindri

code execution

Runs Python, shell, and spreadsheets in an E2B sandbox.

Muninn

long-term memory

Cross-session recall with vault stats, backup, and ChromaDB export.

Norns

thought scorer

Scores thoughts for relevance, evidence, and goal fit.

Forseti

conflict resolver

Conflict resolution with 3-pillar LLM-as-Judge.

Heimdall

insight bridge

Insight bridge between parallel agents.

Gleipnir

loop prevention

Stops loops with breakers and forced convergence.

Valkyrie

early exit & cost

Exits early when paths converge (~2.2× faster), with live USD cost tracking.

Ragnarök

final synthesizer

Final synthesis from the best thoughts across all realms.

Memory & Privacy

You see the reasoning trace. Pick how much Yggdrasil remembers and stores, and use separate profiles (--profile work) to keep 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

Output validation, claim verification, prompt-injection detection, and a first-person crisis guardrail.

Data Controls

Export, download, or clear conversations, memory, and sessions whenever you want.

Private Deployment

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

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

Memory that helps, not hoards

Save & rewind Checkpoint progress, rewind on failure, and retry a different path.
Learn from mistakes Remembers dead ends so the same failure is not repeated.

Research Foundation

Peer-reviewed ideas under the hood, scored on GAIA.

Tree of ThoughtsYao et al., 2023 - multi-path exploration
Graph of ThoughtsBesta et al., 2023 - non-linear aggregation
ReflexionShinn et al., 2023 - self-critique loops
MemGPT / Generative AgentsPacker & Park et al., 2023 - lasting memory

GAIA results

Method notes →

Self-measured on our harness (20 questions per level, 60 total). Engineering evidence, not a leaderboard rank.

Level 1

85%

17/20 correct
~4.6 min · $0.56

Level 2

85%

17/20 correct
~4.4 min · $0.81

Level 3

45%

9/20 correct
~7.0 min · $1.17

Overall

71.7%

43/60 questions
318 min · $50.71 total

Join the project

Core open source planned for
winter 2026

Follow the repository now, then help shape agent architecture, evaluation, project maintenance, documentation, or frontend work when the core opens. Scope and timing may change as the research preview is validated.

Follow on GitHub Connect on LinkedIn

Follow the work now. Contribution guidance will ship with the core.

Join The Research Preview

Private preview access is planned for summer 2026. Accepted participants receive $5 USD in trial credits to test the system and share structured feedback.

Private Research Preview

$5 CREDIT

Included for accepted participants · Summer 2026

Test inspectable parallel reasoning
Compare specialist modes
Share feedback with the creator

More tiers coming soon.

Watch The Work,
Not A Mockup

One system, one memory,
Multiple interfaces

CLI Interface

Chainlit Web UI

REST API & MCP

Bifrost for VS Code & Cursor

Bifrost for Chrome

Open Design & ACP editors

Scheduler

Messaging Soon

One Engine
to Rule Them All

Where Multiple Perspectives Matter

Software Dev

Research

Data Science

Business

Writing / Content

Math & STEM

Document / OCR

Browsing & Shopping

Choose Your Mode

How The System Works

Memory & Privacy

Research Foundation

GAIA results

Core open source planned for
winter 2026

Join The Research Preview

Plant the Seed Today

Custom mode & consultation

Watch The Work,Not A Mockup

One system, one memory,Multiple interfaces

CLI Interface

One Engineto Rule Them All

Where Multiple Perspectives Matter

Software Dev

Research

Data Science

Business

Writing / Content

Math & STEM

Document / OCR

Browsing & Shopping

Choose Your Mode

How The System Works

Memory & Privacy

Parallel AI Reasoning Architecture

Research Foundation

GAIA results

Core open source planned forwinter 2026

Join The Research Preview

Plant the Seed Today

Custom mode & consultation

Watch The Work,
Not A Mockup

One system, one memory,
Multiple interfaces

One Engine
to Rule Them All

Core open source planned for
winter 2026