Lithos¶
The meta system prompt your entire agent team shares¶
Persistent. Searchable. Always up to date.
Lithos is a local, privacy-first knowledge base that lets heterogeneous AI agents — Agent Zero, Claude Code, OpenClaw, LangGraph, CrewAI, and more — read, write, and coordinate through a single MCP interface. Human-readable Markdown on disk. Zero cloud. Zero lock-in.
Around it, three sibling projects feed it, drive it and show it: Influx ingests knowledge, Loom turns tasks into executed work, and Lens is the UI over both.
Markdown-first
Every knowledge item is a plain .md file compatible with Obsidian. Your agents' memory is human-readable and version-controllable — inspect it, edit it, git diff it.
Hybrid search
Tantivy full-text BM25 + ChromaDB semantic vectors, fused with Reciprocal Rank Fusion (RRF) — plus LCMA cognitive retrieval that learns which notes actually help.
Multi-agent coordination
Task claiming with TTL-based locks, plus a full task graph: blocking dependencies, epics, gates on the outside world, and a query for what's ready to work right now.
Knowledge graph
Wiki-links ([[note]]) build a NetworkX graph automatically. Traverse relationships, query provenance lineage, and assert typed edges between notes.
MCP native
Exposes 39 tools via the Model Context Protocol over stdio or HTTP (StreamableHTTP + legacy SSE on one port). Add Lithos to any MCP-compatible agent in seconds — no SDK required.
Truly local
No API keys. No telemetry by default. No cloud sync. Runs on a Raspberry Pi, a Mac Mini, or a VPS. Your data stays where you put it.
Quickstart¶
Lithos is now serving MCP at http://localhost:8765/mcp (StreamableHTTP) and http://localhost:8765/sse (legacy SSE).
Add to your claude_desktop_config.json:
In ~/.openclaw/workspace/config/mcporter.json:
Your agents, talking to each other¶
# Agent A discovers something useful
lithos_write(
title="Rate limiting pattern for OpenAI API",
content="Use exponential backoff with jitter. Base delay 1s, max 60s...",
tags=["openai", "rate-limiting", "patterns"],
agent="research-agent"
)
# Agent B finds it instantly — no re-researching, no duplicated work
results = lithos_search(query="openai rate limit backoff", mode="hybrid")
# → [{ title: "Rate limiting pattern for OpenAI API", score: 0.94, ... }]
# Agent C coordinates parallel work
task = lithos_task_create(title="Audit all API integrations", agent="orchestrator")
lithos_task_claim(task_id=task["task_id"], aspect="OpenAI audit",
agent="agent-c", ttl_minutes=60)
# Agent D just asks what's unblocked and ready to work
ready = lithos_task_ready()
Why Lithos?¶
In 2026, running one agent is table stakes. Running a team of agents is where it gets interesting — and where it gets messy. Agents duplicate research, contradict each other, lose context, and can't coordinate without a shared channel.
Lithos is that channel. It's the shared memory layer your agents can actually trust: every item is timestamped, attributed, versioned, and searchable. Agents can declare confidence, set freshness deadlines, and build provenance chains. The knowledge base is a first-class artefact you can open in Obsidian, commit to git, and inspect at any time.
The AI Runtime
Think of Lithos like a software runtime for knowledge — not a static library. Notes are executable instructions. Outdated notes are bugs. Retrieval feedback adjusts each note's salience over time, and the reconcile pipeline flags stale knowledge.
The Lithos ecosystem¶
Lithos is the hub. The other three projects are separate repositories with their own versions and cadence, documented here under their own tabs.
Lithos · stable, v0.6.0
The knowledge base and MCP server: notes, hybrid search, LCMA retrieval, the knowledge graph, and the task graph agents coordinate through.
Influx · stable, unreleased
Knowledge ingestion: arXiv and RSS sources filtered against interest profiles, LLM enrichment, archiving, and an inbox for URLs and PDFs — all written into Lithos as notes.
Loom · pre-release
Task execution: a daemon that subscribes to Lithos events, claims tasks, and dispatches coding agents through an implement → review → PR pipeline, plus an Obsidian bridge and a GitHub issue watcher.
Lens · pre-release
The web UI: a live task dashboard, dependency graphs, finding timelines, knowledge search and the knowledge graph, with a few curated write actions for an operator.
See how they fit together and the compatibility matrix.