langchain-daftari

The store that knows when it's wrong.

langchain-daftari wraps a Daftari MCP vault as LangChain tools so a LangGraph agent can read, search, write, and curate a long-lived, file-backed knowledge base — instead of re-deriving the same answer every session.

Why

Vector RAG retrieves passages. Daftari stores compiled answers — markdown notes with frontmatter (status, confidence, provenance, decay), git history, and an advisory linter. The vault is the agent's memory across runs.

Plug it into any LangChain/LangGraph workflow and you get four properties for free:

• Search before derive. The wrapper marks vault_search as CRITICAL: Call this BEFORE synthesizing an answer from scratch.
• Long-lived state. Notes persist between sessions. Git history is the audit log.
• Curation. vault_lint flags stale, low-confidence, or unsourced notes.
• Provenance. Every write is auto-committed; vault_provenance traces who added what when.

Install

pip install langchain-daftari
# also install daftari itself (Node.js MCP server)
npm install -g daftari

Requirements:

• Python ≥ 3.10
• Node.js ≥ 18 (Daftari is shipped on npm)

Quick start

from langchain_daftari import DaftariClient, create_daftari_tools
from langgraph.prebuilt import create_react_agent
from langchain_anthropic import ChatAnthropic

with DaftariClient(vault_path="./my-vault", user="me", role="admin") as client:
    tools = create_daftari_tools(client)
    agent = create_react_agent(
        ChatAnthropic(model="claude-sonnet-4-6"),
        tools=tools,
    )
    result = agent.invoke({"messages": [
        ("system", "Search the vault before answering."),
        ("user", "What's our current take on Daftari vs vanilla RAG?"),
    ]})
    print(result["messages"][-1].content)

Don't have a vault yet? Scaffold one with npx daftari --init ./my-vault.

What you get: 14 tools

Tool names, descriptions, and argument schemas come from the live MCP server's tools/list response — never from baked-in copies — so the wrapper layer tracks server changes automatically when you upgrade daftari.

vault_search description override

The wrapper layer prepends one line to vault_search's server-side description:

CRITICAL: Call this BEFORE synthesizing an answer from scratch. The vault may already contain a compiled, reviewed answer.

The override is wrapper-side only — daftari's own description stays neutral so other MCP clients (Claude Code, raw MCP) aren't strong-armed into a LangGraph-specific workflow.

The search-before-derive pattern

Pair the description override with a system prompt that reinforces the discipline:

SEARCH_BEFORE_DERIVE = """\
Core discipline: SEARCH BEFORE YOU DERIVE.

1. Before answering any non-trivial question, call vault_search.
2. If the vault has a compiled note, answer from it and cite the path.
3. If not, do the work, then vault_write a draft so future-you can find it.
4. If a note is stale or incomplete, vault_append or vault_promote.
"""

See examples/demo_research_agent.py for a runnable three-day simulation that asserts the agent actually searches before answering on day 2 and day 3.

Filtering the tool surface

# only read-side tools for a "research-only" agent
tools = create_daftari_tools(client, include={
    "vault_search", "vault_search_related", "vault_read", "vault_status",
})

# everything except destructive curation
tools = create_daftari_tools(client, exclude={"vault_deprecate", "vault_tension_log"})

Architecture

┌─────────────────────────┐
│  LangGraph ReAct agent  │
└────────────┬────────────┘
             │  StructuredTool.invoke / ainvoke
             ▼
┌─────────────────────────┐
│  create_daftari_tools   │   builds one StructuredTool per MCP tool
└────────────┬────────────┘
             │  DaftariClient.call_tool / acall_tool
             ▼
┌─────────────────────────┐
│      DaftariClient      │   subprocess + ClientSession on a
│   (transport primitive) │   dedicated background event loop
└────────────┬────────────┘
             │  JSON-RPC over stdio (MCP)
             ▼
┌─────────────────────────┐
│  daftari MCP server     │   Node.js, manages vault + SQLite index + git
└─────────────────────────┘

DaftariClient runs the MCP session on its own background event loop so a single client can be shared safely by multi-threaded sync callers and async event-loop callers without each call having to spawn a fresh subprocess.

Reference: DaftariClient

DaftariClient(
    *,
    vault_path: str,                          # required
    user: str = "guest",
    role: str = "guest",
    command: list[str] | None = None,         # default: ["npx", "daftari"]
    env: dict[str, str] | None = None,
    timeout: float = 30.0,
)

• Pass command=["daftari"] for a global npm install.
• Pass command=["node", "/path/to/daftari/dist/cli.js"] to run a local clone.
• Pass command=["npx", "-y", "daftari@1.12.6"] to pin a version.

Sync surface: client.call_tool(name, args) returns DaftariResponse. Async surface: await client.acall_tool(name, args) does the same, safe from any event loop.

DaftariResponse has:

• .text — concatenated text content blocks
• .data — parsed JSON if the text looks like JSON, otherwise the raw string
• .is_error — whether the MCP server flagged the call as an error
• .raw — the underlying mcp.types.CallToolResult for advanced inspection

Compatibility

langchain-daftari is compatible with daftari ≥ 1.12.0, < 2.0.0 on npm.

The compatibility line is documented here rather than pinned as a Python dependency because daftari ships as an npm package, not a Python package. The package will get a major version bump on the Python side if any of the following happens server-side:

• A tool is removed.
• A tool's input schema changes in a breaking way.
• The MCP protocol version changes.

Tool additions and non-breaking schema changes do not require a major bump because the wrapper layer reads schemas live from tools/list.

Development

cd integrations/langchain
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"

pytest                            # all 34 tests (28 mock + 6 integration)
pytest -m "not integration"       # mock-only, no Node.js required

The integration tests boot a real daftari subprocess via npx. They skip themselves cleanly if npx isn't on PATH.

Status & roadmap

This is phase 1 — a thin LangChain tool wrapper over the daftari MCP surface. DaftariClient is deliberately LangChain-free so the same transport primitive can serve future integrations.

What's next — read-side LangMem audit: The decided direction for phase 2 is a read-side audit layer that productizes the existing langgraph-store-demo. Rather than acting as a write-back store, daftari reads whatever store LangMem already writes (e.g. Postgres), imports those memories read-only, runs tension detection against the vault, and compiles claim notes. BaseStore's flat put() seam can't carry daftari's compiled-note / provenance / tension value at write time, so being a store backend would deliver none of daftari's differentiation.

Out of scope for this release: an async-first user surface, LangServe deployers, and any opinionated retriever / chain abstractions on top of the raw tools.

License

MIT. See LICENSE.