kb0-mcp 0.3.0

Python client for kb0 — the knowledge base layer for AI agents (MCP, markdown, git).

kb0 (Python client)

Native Python client for kb0 — the knowledge base layer for AI agents. Markdown-first, git-backed, MCP-native.

This package hides the MCP plumbing and gives you a clean async VaultClient.

pip install kb0-mcp

The PyPI package is kb0-mcp; you still import kb0 in code.

Requires the kb0 server. This is a client — it spawns the kb0 binary as a subprocess. Install it once with Node: npm install -g kb0-mcp, then kb0 init my-vault.

Quickstart

import asyncio
from kb0 import VaultClient

async def main():
    async with VaultClient(vault="./my-vault", agent="my-bot") as kb:
        # write a note (author, id, timestamps set by the server)
        res = await kb.write(
            "notes/auth.md",
            title="Auth: we chose JWT",
            content="Short-lived access + refresh tokens. See [[notes/security.md]].",
            tags=["security"],
        )
        print("created", res["path"], "hash", res["hash"])

        # search (hybrid by default)
        hits = await kb.search("authentication design")
        for r in hits["results"]:
            print(r["score"], r["title"], r["path"])

        # read + update with optimistic locking
        note = await kb.read("notes/auth.md")
        await kb.update(
            "notes/auth.md",
            content=note["content"] + "\n\nUpdate: rotating keys monthly.",
            expected_hash=note["hash"],
        )

asyncio.run(main())

Hosted vault (kb0 cloud)

Point the same client at a kb0:// address with your API key and the vault lives on the kb0 cloud — no local files, no kb0 engine install, pure Python over HTTPS. Same methods; every operation lands in your audit trail:

async with VaultClient(
    vault="kb0://team-kb",
    agent="my-bot",
    api_key="kb0_live_...",   # create one in the kb0 dashboard (or set KB0_API_KEY)
) as kb:
    await kb.write("notes/idea.md", title="Idea", content="...")
    hits = await kb.search("idea")          # hosted keyword search
    graph = await kb.backlinks("notes/idea.md")

On a local vault, passing api_key= (or setting KB0_API_KEY) enables content-free audit forwarding to the same dashboard.

Configuration

VaultClient passes everything through to kb0 serve:

VaultClient(
    vault="./my-vault",       # vault directory
    agent="my-bot",           # agent identity (provenance + ACL)
    openai_api_key="sk-...",  # optional — enables semantic search
    strict=False,             # require .vault-policy.yaml if True
    command="kb0",            # override the binary path if needed
    env={"KB0_EMBEDDING_MODEL": "text-embedding-3-large"},
    api_key="kb0_live_...",   # optional — audit forwarding / hosted vaults
    cloud_url=None,           # override the kb0 cloud base URL (kb0:// vaults)
)

API

All methods are async and return plain dicts (the tool's structured output).

Method	kb0 tool
await kb.write(path, *, title, content, status="draft", tags=None)	vault.write
await kb.read(path)	vault.read
await kb.update(path, *, content, expected_hash, title=None, status=None, tags=None)	vault.update
await kb.delete(path)	vault.delete
await kb.search(query, *, mode="hybrid", ranking="rrf", limit=10, filters=None)	vault.search
await kb.list(*, prefix=None, tag=None, status=None, limit=50)	vault.list
await kb.recent(limit=10)	vault.recent
await kb.backlinks(path)	vault.backlinks
await kb.links(path)	vault.links
await kb.status()	vault.status

Errors

Failures raise typed exceptions you can catch:

from kb0 import KbConflictError, KbACLDeniedError, KbNotFoundError

try:
    await kb.update("notes/x.md", content="...", expected_hash=stale_hash)
except KbConflictError:
    note = await kb.read("notes/x.md")   # re-read, get the current hash, retry

KbError is the base class; KbNotFoundError, KbConflictError, KbValidationError, and KbACLDeniedError are its subclasses.

Why a subprocess?

kb0 is an MCP server. The Python client launches kb0 serve over stdio and speaks MCP to it — the same protocol Claude Desktop uses. Your vault stays a local folder of markdown under git; this client is just an ergonomic way for Python agents to talk to it. See the main repo for the architecture.

Audit log

Every call you make through the client is recorded server-side to <vault>/.vault-index/kb0.log as content-free JSON lines — the path read, the query searched, and the paths a search returned, never the note bodies. Failed and denied calls are logged too, so you can audit what an agent read and searched, not just what it changed (git already covers writes).

License

Apache 2.0.