Skip to main content

Coming soon

Reason this release was yanked:

dev

Project description

mcp_context_pointer

A plug-in extension for Model Context Protocol (MCP) enabling Cross-App Memory Pointers (ContextPointer block) and seamless use of vaulted memory with LLM agents.
Built by Third Plane.

PyPI version MCP compatibility License: Apache-2.0


🚀 What is this?

mcp_context_pointer is an optional plug-in for the MCP Python SDK.
It provides the ContextPointer block—an efficient, privacy-stable, cross-application pointer for referencing conversation history, memory chunks, or arbitrary state, backed by a vault service.

  • Save up to 40–60% on LLM context tokens
  • Enable compliance (GDPR, Consent Ledger) and auditability
  • Point to fast, managed vaults or your own self-hosted instance

Supports:

  • Adaptive inlining out-of-the-box (fallbacks when vault or network is slow)
  • Signature verification (Ed25519/PASETO or HMAC)
  • Works with any MCP-compliant service, agent, or app

🛠️ Installation

pip install mcp-context-pointer
# or as an MCP extra:
pip install "modelcontextprotocol[context-pointer]"

💡 Quick Start

from mcp_context_pointer.block import ContextPointerBlock
from mcp.blocks.registry import get_block
import requests

# Compose a conversation:
history = [
    {"role": "user", "text": "Can you summarize our last meeting?"},
    {"role": "assistant", "text": "Certainly, here's a short summary..."},
    {"role": "user", "text": "What was the action item?"}
]

# Inline last N, pointerize the rest:
from mcp_context_pointer.wrap import wrap_history_with_pointer

new_blocks = wrap_history_with_pointer(
    history=history,
    n_inline=2,
    vault_url="https://vault.thirdplane.io"          # managed or self-host
)

# Send new_blocks as your MCP resource.blocks[]

🏛️ ContextPointer Block Spec (v1)

{
  "kind": "context_pointer",
  "mrl":  "ocsv1://vault.thirdplane.io/u/1234/c/77#t0-4",
  "sig":  "{signature}",
  "privacy": "app_private",
  "window": {"start": 0, "end": 256},
  "maxCost": 512
}
  • mrl: Opaque, signed Memory Reference Locator. Supports privacy, region pinning, selectors.
  • sig: Signature (Ed25519/PASETO default, HMAC fallback)
  • privacy: Recommended: 'app_private' (default: cross-app only if user explicit)
  • window: Optional byte or token range
  • maxCost: LLM context size hint

🔒 Security

  • All pointers must be signed and verified on-dereference.
  • privacy flags default to non-leaky mode; PII should never appear in mrl.
  • Vault fetch enforces same-site/tenant and access logging.
  • Supports Ed25519 signatures by default; accepts legacy HMAC.

🏎️ Performance

  • Fetch pointer blocks from edge-cached managed vaults in <35ms (p95, global).
  • Fallback: If vault unreachable, package will inline fallback for full compatibility ("never worse" policy).
  • Prometheus (or JSON) metrics on all pointer ops.

📝 License

Apache-2.0 — matches MCP ecosystem, allows both independent & up-streaming use, and is enterprise-friendly.
Copyright (c)


📡 Roadmap

  • 🚀 Upstream into MCP core once the ContextPointer profile is standardized
  • Edge cache + multi-region support for managed vault
  • Optional Consent Ledger & Answer-Cache APIs
  • Thin wrappers for Node, Go, Rust

🤝 Acknowledgments

Special thanks to the MCP community, spec maintainers, and early design partners. Questions? Open an issue or DM .


⭐ About Third Plane

Third Plane is building the "Cloudflare for AI State"—portable, privacy-stable, edge-cached memory for every AI app.


Legal

This package is independent of, but interoperates with, the core MCP SDK.
No license or support is implied by the Model Context Protocol working group.


Start saving tokens and gaining memory: Read the full docs ➔ or Try managed vault for free ➔


FAQ

Q: Can I use this without a managed vault?
A: Yes—just set vault_url to your own self-hosted endpoint.

Q: How do I upstream to MCP if the spec merges?
A: Final API/class names will be stable; your wrap() just imports from MCP core once standardized.

Q: Does this lock me in?
A: No. It is Apache-2.0, and all data is exportable via MRL.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

mcp-context-pointer-0.0.1.tar.gz (4.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

mcp_context_pointer-0.0.1-py3-none-any.whl (4.1 kB view details)

Uploaded Python 3

File details

Details for the file mcp-context-pointer-0.0.1.tar.gz.

File metadata

  • Download URL: mcp-context-pointer-0.0.1.tar.gz
  • Upload date:
  • Size: 4.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.1.0 CPython/3.10.16

File hashes

Hashes for mcp-context-pointer-0.0.1.tar.gz
Algorithm Hash digest
SHA256 fdbd477465ee7da4910725a1149f23d2e8ceea293f94c783f7731495c1fb3051
MD5 1e0bea14c2de3e707bba3a8b2fe6c6fa
BLAKE2b-256 dd6f2f7695ff8a365aafc055c8b201135534f8f809387bcee14cdf639c9490aa

See more details on using hashes here.

File details

Details for the file mcp_context_pointer-0.0.1-py3-none-any.whl.

File metadata

File hashes

Hashes for mcp_context_pointer-0.0.1-py3-none-any.whl
Algorithm Hash digest
SHA256 3f937594eb644f489cca89219775dc0b658033e4a47ed7e7919343bdcbdfa98d
MD5 0834173b36eb11ec3cee6fceed581d5c
BLAKE2b-256 38d7542f8cde47b939616f1e2ff4f7db4f0f5af2fab5f278fd907c82091afcfd

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page