Use Claude Code or your agents to run projects, not just code: it stays on track and learns from your work, no slip-ups.
Sarah is your colleague who knows the why behind every decision — and which 3% of the day is worth writing down. Inside the agent, no memory. Around the agent, no filter. Sarah does both. Your stack does neither.
The agent re-derives your constraints, re-discovers conventions, re-invents the wheel each chat. Output slowly diverges from intent — a different naming scheme here, a contradicted decision there, a forgotten guardrail somewhere else.
Day one looks fine. Day ninety, none of it traces back to you.
Notion, Obsidian, a Confluence wiki, a CLAUDE.md per repo — every tool you reach for is a bucket, not a brain. They keep whatever gets dumped in. None of them discard the 97% noise to surface the 3% that matters.
Captures decisions as agents work. Structures them as a cross-project graph. Hands them back to every agent — as MCP skills they call.
Decisions, specs, playbooks, learnings, errors — captured as a side-effect of the work your agents already do. No human-in-the-loop typing.
Every agent output is scored, deduplicated, and linked. The 97% noise stays in the buffer. The 3% that matters becomes a typed node, connected to past decisions across every project.
Your agents stop guessing. They query the graph directly. Blast radius. Impact. Cited by. Depends on. Answered in milliseconds, before the next prompt.
You don't visit a dashboard. You don't migrate data by hand. Your Claude Code runs the wizard, and Marvis lives inside the stack you already use.
From inside Claude Code, Codex, or Cursor — the same terminal you already work in.
$ claude run "install marvis from justaskmarvis.com"
Folder scaffold, governance hooks, MCP server. The agent updates its own CLAUDE.md / AGENTS.md so future sessions wake up brain-aware.
→ cloning marvis-harness … → updating CLAUDE.md, AGENTS.md → installing 5 governance hooks → wiring MCP server at .marvis/mcp.sock ✓ harness in place
Existing repos, decisions, READMEs, agent transcripts — indexed into the knowledge graph. Constitution applied across every project at once. No manual migration.
→ scanning ~/code · found 12 projects → ingesting · 1,847 nodes · 4,212 edges → constitution applied to 12/12 ✓ brain online
Every session reads the brain. Every result writes back, tagged with your approval. Each morning a digest tells you what the brain has noticed since you left.
$ claude › Brain: 3 unresolved decisions from last week. › Brain: onboarding agent broke a rule — flagged. › Brain: new learning extracted from 4 converged sessions. › _
parseUserAgent fix. Promote to playbook?enterprise-bender — 3 days idle.Five categories people recognise on first glance. Marvis takes each one further.
Marvis is open source forever. MarvisX is what you pay for when you need departments, sovereignty, or hands off the keyboard.
Self-host. Single-user. Terminal-first.
Multi-user. Sovereign data. Whitelabel.
For those who don't live in the terminal.
For companies that want AI to ship work — without giving up uptime, departmental separation, or where the data lives.
Product knows what Sales decided. Engineering sees Compliance constraints. Departments share context without sharing everything. SSO, SCIM, role-based scopes.
Your team's agents run server-side, not on laptops. No "my Mac died at 3am" stories. Reachable from anywhere on day one, audit-logged everywhere.
Data stays on your servers, never ours. Run an in-house LLM, or BYOK against OpenAI, Anthropic, Mistral. No vendor lock-in on infra or model.
Every action chains back to the human who said yes. Logging and human-oversight requirements on by default. Your legal team writes policy; MarvisX enforces it.
Things people ask before installing.
You don't change tools. You don't migrate. You drop in one line and your stack wakes up with a brain.