AgentCRDT
Provides OpenAI-compatible tool definitions for managing world facts and detecting contradictions, enabling integration with OpenAI's chat completions API.
agentcrdt
Semantic-causal CRDT for agent-mutable world state.

Quick Start · How It Works · CLI Reference · MCP / Claude · OpenAI · vs. Alternatives · Contributing
Why
Multi-agent systems share a mutable view of the world. When two agents update the same fact simultaneously — or when one agent asserts a fact that logically contradicts another — the shared state becomes inconsistent.
Traditional approaches either:
Lock everything — too slow for real-time agents
Last-write wins blindly — loses causal context and semantic validity
Ignore contradictions — downstream agents act on stale or conflicting beliefs
agentcrdt solves this with a semantic-causal CRDT:
Every fact is content-addressed by
(domain, entity, attribute)— same key always resolves to the same slotConcurrent updates merge via Last-Write-Wins (version then timestamp)
Semantic rules detect logical contradictions across domains (e.g. a dead king can't have a valid treaty)
Contradiction events are persisted as first-class objects for audit and remediation
agentcrdt set life king alive False --agent-id agent-A
agentcrdt merge agent-B.db
agentcrdt events # shows ContradictionEvents if rules are violatedRelated MCP server: iranti
How It Works
flowchart LR
A[Agent A\nWorldFact version=2] --> M[WorldMerger\nLWW merge]
B[Agent B\nWorldFact version=1] --> M
M --> S[WorldStore\nSQLite]
S --> R[RuleEngine\nsemantic check]
R --> C[ContradictionEvent\nsaved to store]
R --> OK[No event\nconsistent state]Core primitives:
WorldFact — content-addressed by
SHA-256[:16]("domain|entity|attribute"). Same key across agents = same slot. Value and version are mutable metadata.SemanticRule — a first-order implication: "if
domain.entity.attr=Xthenother_domain.entity.other_attrmust beY".RuleEngine — evaluates rules against the merged world state. Returns
ContradictionEventobjects for violations.WorldMerger — merges two stores via LWW, then runs the rule engine.
WorldStore — SQLite-backed fact and event store. Thread-unsafe; one store per process.
Features
Feature | Details |
Content-addressed facts | Same |
LWW CRDT merge | Higher version wins; on tie, higher timestamp wins |
Semantic rules |
|
Contradiction events | Persisted as |
Offline / local-first | Single SQLite file, no server required |
JSON output | Machine-readable output for downstream automation |
Markdown output | Ready-to-paste GitHub PR comment |
FastAPI REST server |
|
MCP server | Model Context Protocol integration for Claude and other agents |
52 tests | Comprehensive test suite covering all layers |
Quick Start
pip install agentcrdtfrom agentcrdt import WorldFact, WorldStore, WorldMerger, SemanticRule, RuleEngine
# Agent A: king is dead
fact_a = WorldFact(domain="life", entity="king", attribute="alive",
value=False, version=1, agent_id="agent-A")
# Agent B: treaty is still valid (inconsistent!)
fact_b = WorldFact(domain="alliance", entity="king", attribute="valid",
value=True, version=1, agent_id="agent-B")
rule = SemanticRule(
name="dead-king-voids-treaty",
trigger_domain="life", trigger_attribute="alive", trigger_value=False,
implies_domain="alliance", implies_entity_same=True,
implies_attribute="valid", implies_value=False,
)
with WorldStore("local.db") as local, WorldStore("remote.db") as remote:
local.set_fact(fact_a)
remote.set_fact(fact_b)
result = WorldMerger(rule_engine=RuleEngine([rule])).merge(local, remote)
print(result.conflicts) # [ContradictionEvent(rule='dead-king-voids-treaty', ...)]
# Note: the snippet above creates local.db and remote.db in the cwd; delete them when done.
# For automatic cleanup wrap with tempfile.TemporaryDirectory() (see docs/quickstart.md).CLI Reference
agentcrdt [--db PATH] COMMAND [ARGS]Command | Description |
| Store or update a world fact |
| Retrieve a world fact by key |
| Merge another store into this one (LWW CRDT) |
| List all contradiction events |
| Show fact count and event count |
| Show installed version |
Examples:
agentcrdt --db world.db set life king alive False --version 2 --agent-id agent-A
agentcrdt --db world.db get life king alive
agentcrdt --db world.db merge /tmp/agent-b.db
agentcrdt --db world.db events
agentcrdt --db world.db statusREST Server
pip install 'agentcrdt[api]'
uvicorn agentcrdt.api:app --reloadEndpoints:
Method | Path | Description |
|
| Create or update a world fact |
|
| List facts (optional |
|
| Merge a remote store into the local one |
|
| List contradiction events |
|
| Health check |
MCP / Claude Desktop Integration
pip install 'agentcrdt[mcp]'Add to ~/.config/claude/claude_desktop_config.json:
{
"mcpServers": {
"agentcrdt": {
"command": "agentcrdt-mcp"
}
}
}Available MCP tools: set_world_fact, get_world_facts, merge_world_state.
OpenAI Integration
agentcrdt exposes OpenAI-compatible tool definitions in tools/openai-tools.json:
import json
tools = json.loads(open("tools/openai-tools.json").read())
# Pass `tools` to openai.chat.completions.create(tools=tools, ...)Topics: #crdt #agents #multi-agent #world-state #llmops #mcp
Python API
src/agentcrdt/
├── fact.py # WorldFact, ContradictionEvent
├── rules.py # SemanticRule, RuleEngine
├── store.py # WorldStore (SQLite)
├── merger.py # WorldMerger, MergeResult
├── report.py # print_state, print_events, to_json, to_markdown
├── cli.py # Click CLI entry point
├── api.py # FastAPI REST server
└── mcp_server.py # MCP serverReal-World Scenario
E-Commerce: Preventing Multi-Agent Inventory Oversell on Black Friday
Three purchasing agents — East Coast, West Coast, and EU — each read inventory=50 and begin processing orders simultaneously. Without coordination, they'd oversell by 25 units ($380K in chargebacks). With agentcrdt, the merge surfaces the contradiction immediately:
import tempfile
import os
from agentcrdt import WorldFact, WorldStore, WorldMerger, SemanticRule, RuleEngine
# SemanticRule: if inventory count goes negative (EU sold past zero),
# the oversell_guard must flip to True — but the system seeded it False.
# The mismatch fires a ContradictionEvent at merge time.
oversell_rule = SemanticRule(
name="inventory-oversell-detected",
trigger_domain="inventory",
trigger_attribute="count",
trigger_value=-25, # EU's final count after selling 45 of 50
implies_domain="inventory",
implies_entity_same=True,
implies_attribute="oversell_guard",
implies_value=True, # rule expects guard=True when count<0
)
with tempfile.TemporaryDirectory() as tmp:
east_db = os.path.join(tmp, "east.db")
west_db = os.path.join(tmp, "west.db")
eu_db = os.path.join(tmp, "eu.db")
local_db = os.path.join(tmp, "local.db")
# System seeds a guard fact: "no oversell in progress" (version=0)
with WorldStore(local_db) as local:
local.set_fact(WorldFact(
domain="inventory", entity="airpods_pro",
attribute="oversell_guard", value=False,
version=0, agent_id="system",
))
# All three agents read inventory=50 simultaneously and start selling.
# East sells 30 → remaining count = 20
with WorldStore(east_db) as east:
east.set_fact(WorldFact(
domain="inventory", entity="airpods_pro",
attribute="count", value=20,
version=1, agent_id="east",
))
# West sells 25 → remaining count = 5 (already below East's view)
with WorldStore(west_db) as west:
west.set_fact(WorldFact(
domain="inventory", entity="airpods_pro",
attribute="count", value=5,
version=1, agent_id="west",
))
# EU sells 45 → count = -25 (went deeply negative)
with WorldStore(eu_db) as eu:
eu.set_fact(WorldFact(
domain="inventory", entity="airpods_pro",
attribute="count", value=-25,
version=2, agent_id="eu",
))
# Merge all three into local — LWW picks EU's version=2 count=-25.
# The rule engine then sees count=-25 alongside oversell_guard=False
# and fires a ContradictionEvent.
merger = WorldMerger(rule_engine=RuleEngine([oversell_rule]))
with WorldStore(local_db) as local, \
WorldStore(east_db) as east, \
WorldStore(west_db) as west, \
WorldStore(eu_db) as eu:
merger.merge(local, east)
merger.merge(local, west)
result = merger.merge(local, eu)
print("Conflicts detected:", len(result.conflicts))
for evt in result.conflicts:
print(f" rule={evt.rule!r} agents={evt.agent_a!r} vs {evt.agent_b!r}")
# Conflicts detected: 1
# rule='inventory-oversell-detected' agents='eu' vs 'system'What this prevents: Last-write-wins databases silently allow the oversell — the last agent to write wins and the inventory appears valid until chargebacks arrive. agentcrdt surfaces the semantic contradiction at merge time so a reconciliation agent can intervene before fulfillment.
vs. Alternatives
Tool | Approach | Semantic Rules | MCP |
agentcrdt | Content-addressed LWW CRDT | Yes | Yes |
Automerge | JSON CRDT | No | No |
Yjs | CRDT text/map | No | No |
Redis | Key-value store | No | No |
custom DB | Ad-hoc | Manual | No |
Star History
Smithery
agentcrdt is available on Smithery — search for agentcrdt.
Case Studies
See how teams are using agentcrdt in production:
Conflict-Free World State for 500k Concurrent MMO Players — Cascade Games eliminates boss-fight HP flickering and enables 500k concurrent players
Conflict-Free Shared Knowledge Base for a 6-Agent Research Pipeline — Cognition Labs reduces report synthesis from 3 hours to 12 minutes
Stay Updated
Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.
Contributing
See CONTRIBUTING.md. PRs reviewed within 5 business days.
Closed loop / Non-Ornament
See docs/CLOSED_LOOP.md for when this library is load-bearing vs ornamental, and when not to use it.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- AlicenseAqualityCmaintenanceCollective memory for AI agents. One agent solves a bug - every agent in the world gets the fix instantly.3MIT
- AlicenseAqualityAmaintenancePersistent shared memory for AI coding agents. Stores facts as entity/key/value triples with hybrid semantic search, task checkpoints, and conflict resolution — shared across Claude Code, Codex CLI, and GitHub Copilot.162355AGPL 3.0

cafecitoofficial
Alicense-qualityBmaintenanceEnables AI agents to coordinate on a shared repository using commutativity-proven parallel landing and regenerative merge, avoiding rebase conflicts through symbol-level leases.Apache 2.0- Alicense-qualityBmaintenanceHeadless, peer-to-peer context synchronization for local AI agents. It enables multiple LLM agents to share structured context and resolve state conflicts over a serverless P2P network.2MIT
Related MCP Connectors
Collective memory for AI agents. One agent solves a bug — every agent gets the fix instantly.
Long-term memory for AI agents: semantic facts, episodic events, and procedural workflows
Shared, verifiable memory for AI agents and robots: signed tokens that resolve and verify offline.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/sandeep-alluru/agentcrdt'
If you have feedback or need assistance with the MCP directory API, please join our Discord server