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 slot
• Concurrent 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 violated

---

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=X then other_domain.entity.other_attr must be Y".
• RuleEngine — evaluates rules against the merged world state. Returns ContradictionEvent objects 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 (domain, entity, attribute) always produces the same ID
LWW CRDT merge	Higher version wins; on tie, higher timestamp wins
Semantic rules	SemanticRule + RuleEngine detect cross-domain contradictions
Contradiction events	Persisted as ContradictionEvent objects in the store
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	/fact, /facts, /merge, /events, /health endpoints
MCP server	Model Context Protocol integration for Claude and other agents
52 tests	Comprehensive test suite covering all layers

---

Quick Start

pip install agentcrdt

from 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
set DOMAIN ENTITY ATTR VALUE	Store or update a world fact
get DOMAIN ENTITY ATTR	Retrieve a world fact by key
merge OTHER_DB	Merge another store into this one (LWW CRDT)
events	List all contradiction events
status	Show fact count and event count
--version	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 status

---

REST Server

pip install 'agentcrdt[api]'
uvicorn agentcrdt.api:app --reload

Endpoints:

Method	Path	Description
POST	/fact	Create or update a world fact
GET	/facts	List facts (optional ?domain= filter)
POST	/merge	Merge a remote store into the local one
GET	/events	List contradiction events
GET	/health	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 server

---

Real-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.