Atlas is a self-hosted knowledge platform for people and agents.
People use Atlas to organize governed documents, collaborate in Notes, and ask questions with inspectable evidence. Agents use Atlas through MCP to research governed Project knowledge and read the evidence behind their results.
Atlas is more than a document-chat or RAG interface: it brings human and agent access, authorization, and evidence inspection into the same knowledge workflow.
Current public snapshot: local technical preview ·
resettable_developmentdata lifecycle · not Release Ready or Internet Ready
Start Atlas locally · Read the documentation · Join the discussion
The Atlas Web workspace after guided setup—the human entry point for governed Projects, Teams, documents, Notes, conversations, and administration. Provider credentials and sample knowledge are not included in this public snapshot.
- One governed knowledge platform: People work through the Web UI while Agents research governed Project knowledge through MCP. Documents, Notes, identity, model routing, administration, and Agent access run in one self-hosted deployment.
- Authorization stays current: Atlas evaluates access when knowledge is used and rechecks protected evidence reads before returning content.
- Results stay inspectable: Human answers and Agent research expose evidence and auditable execution state instead of acting like opaque model output.
| People | Agents | |
|---|---|---|
| Entry point | Web workspace | MCP |
| Knowledge access | Project and Team knowledge | Project knowledge |
| Primary work | Manage documents and Notes; ask through General or In-depth | Discover scopes, run one-round research, retrieve results, and inspect evidence |
| Result | Answer, execution state, and inspectable evidence | Research packet, optional grounded answer, and inspectable evidence |
Agent access is intentionally narrower than the human workspace in this public snapshot. Agents do not upload documents or edit Notes through MCP.
People ──→ Web workspace ──→ Project / Team knowledge ──→ Answers + evidence
Agents ──→ MCP research ───→ Project knowledge ──→ Packets + evidence
Human conversations can use Project or Team knowledge, while MCP Agents are limited to Project scopes granted to their identity. Both paths expose inspectable evidence and auditable results through the same Atlas deployment.
See Architecture and trust boundaries for exact authority, access, evidence, failure, and lifecycle behavior.
You need Git, Docker Engine with Docker Compose v2, enough memory and disk for the complete stack and pinned embedding-model cache, and a fresh disposable evaluation environment.
git clone https://github.com/redstone39/Atlas.git
cd Atlas
docker compose up --build -dNo .env file is required for a fresh local evaluation. The root compose.yml
starts the complete local stack and creates persistent deployment secrets when
explicit overrides are absent.
Open http://127.0.0.1:5184/login. An empty deployment redirects to the resumable guided setup:
Administrator → Model → Project → Document → Review
Model, Project, and Document may be skipped, but a first governed answer needs a tested default model route, accessible Project knowledge, and searchable evidence.
For health checks, recovery, replacement, and lifecycle details, see
Local Docker Compose deployment. Its documented
down -v reset permanently deletes local Atlas application data. For secrets,
Providers, MCP, and runtime inputs, see
Configuration.
- Add documents to a Project or Team and wait until they are Searchable.
- Create or collaborate in scope-bound Notes.
- Ask a question through a General or In-depth conversation.
- Inspect the answer, execution state, and available evidence.
Evidence review helps inspect what Atlas used; it is not a truth or formal citation guarantee. Model answers are not deterministic.
A System Admin can create an Agent identity, issue its opaque bearer token, and grant it Project access.
- Discover available Project scopes.
- Submit one bounded, single-round research request.
- Retrieve the research packet.
- Inspect selected evidence.
An MCP 2025-06-18 client connects to exact /mcp. The research packet is the
primary result, and an optional governed answer remains bound to it. Fresh
requests and protected evidence reads re-evaluate access.
- Fresh local Docker Compose deployment with loopback Web and MCP access.
- Governed documents, collaborative Notes, scoped conversations, and inspectable evidence.
- Local Provider and model configuration.
- One-round MCP Agent research over Project knowledge.
- No direct public Internet deployment.
- No in-place application-data migration; replacing a snapshot requires a fresh application data set.
- No high availability, automatic failover, or shared-state multi-deployment.
- No managed TLS, backup, monitoring, capacity management, abuse protection, or operations SLA.
Keep the default Compose ports bound to loopback. The default MCP transport accepts localhost Host and Origin values only; configuring an operator-managed proxy or public URL does not make this snapshot Internet Ready.
LDAP and Active Directory ship unconfigured. Portainer with SMB paths are documented and audited, but real-environment operation is not verified in this public snapshot.
These boundaries are open to discussion. If Atlas is missing something important for your use case, share the scenario, expected outcome, and relevant environment details through Issues or Discussions.
System Admin manages identity, Projects and Teams, documents, models, processing, agents, audits, system status, language, theme, Notes settings, and guided setup. The pictured account is local evaluation data.
| Your question | Start here |
|---|---|
| How does Atlas work and who owns each decision? | Architecture and trust boundaries |
| How do first-admin setup, secrets, Providers, directory identity, MCP, and runtime limits work? | Configuration |
| How do I operate or reset a fresh local deployment? | Local deployment |
| How do I verify the public source tree? | Verification |
| How can I report feedback or discuss Atlas? | Community and feedback |
Atlas is also a reference workload for a harness-engineered development methodology: humans retain ownership of product intent, architecture, risk, and acceptance, while bounded coding agents work against explicit repository contracts and verification. See Development methodology for details.
Atlas is feedback-open and discussion-open. External pull requests and unsolicited patches are not currently accepted.
Use GitHub Issues for specific, reproducible defects and reports. Use GitHub Discussions for questions, experiences, architecture discussion, and broader ideas. See CONTRIBUTING.md for the complete participation boundary.
Do not publish vulnerabilities, credentials, private documents, exploit details, or unredacted logs; follow the private-reporting direction in SECURITY.md. No support SLA or implementation timeline is offered.
Atlas source in this repository is licensed under Apache License 2.0. See LICENSE, NOTICE, and THIRD_PARTY_NOTICES.md.