Skip to main content
Pre-release

This release is a pre-release and may not be stable for production use.

phiward

phiward keeps protected data (for example patient data) and your credentials out of what an AI coding agent sends to its model provider. The agent's tools send their output to the provider, and that output can hold file contents, sample identifiers or file names. With phiward, the agent works on protected data only through a gateway that masks names and identifiers.

Why it exists, and how it differs from sandboxes, containers and PII tools: docs/why.md.

Start here: docs/quickstart.md. Three steps: install phiward, set it up (with Claude Code's help or by hand), and check it. In short:

uv tool install --python-preference only-managed phiward==0.2.0b1

then in Claude Code, /plugin marketplace add msk-access/phiward#v0.2.0b1, /plugin install phiward@phiward and /phiward:setup.

Status: v0.2.0b1, a beta. Tested with the real Claude Code on macOS, in the desktop app and in the terminal, and against a real HPC cluster. On Linux, only the test suite has run (no real Claude Code session yet). Read docs/threat-model.md before you rely on it for real data.

It is a Python package with no dependencies. It makes no network connection of its own and has no telemetry. It needs Python 3.9 or newer, macOS or Linux, and Claude Code 2.1.295 or newer.

How it works

  The AI agent's tools                                What reaches the model provider
  (Read, Grep, Bash, MCP, ...)
          |
          |  not protected (the project, your prompts)
          +---------------------------------------->  as before, unchanged
          |
          |  a protected folder or a credential?
          v
  phiward's hook + the sandbox ---- refused ------->  only the refusal and its reason
          |
          |  work on protected data
          v
  The gateway (phiward run / copy)
    a broker outside the agent runs the command
    in a kernel sandbox, then masks the output  ---->  the command's output, names and
                                                       identifiers masked (so print
                                                       counts, not records)
  • Two modes. Mode A, the default: a normal Claude Code session (the desktop app, an IDE, a terminal), wired by phiward install claude. Claude Code's own sandbox confines Bash; the file tools are refused protected data and credentials by phiward's hook and by permission rules. Mode B, the stronger one: phiward launch runs the whole agent tool in phiward's kernel profile (macOS Seatbelt, Linux bubblewrap). Terminal only for now. See docs/modes.md.
  • The gateway. The agent works on protected data only through phiward run -- CMD and phiward copy. The broker runs them outside the agent, as you, and masks the output: protected paths, file names and identifiers become placeholders. HPC goes through phiward run --remote and phiward copy, over your own shared SSH connection (docs/hpc.md).
  • Git hooks refuse commits that hold protected names or identifiers.

What it does not do

  • Masking hides names and identifiers, not other content: through the gateway, print counts and summaries, not records.
  • Out of the box, masking inside file contents covers only sequence and alignment file names, read names and loci (the genomics pack). Subject and sample identifiers inside files are masked only with a pack that has them (the msk example), your own [sanitize] patterns or names_from. The gateway masks entry names under the protected roots (4 characters or longer). Planned for 0.2.0b2: setup asks, doctor warns, and the output says what was masked (#3, #5).
  • In Mode A, the file tools are refused inside Claude Code, not by the kernel; hooks and MCP servers that are not wrapped run with your rights; credentials that are not files (the macOS Keychain, some Kerberos caches, tokens in your environment) are not hidden by a file list.
  • The network is open by default: an agent can send what it can read. In Mode A, [network] limit (suggested by phiward setup) limits shell commands to listed hosts, so git push and pip install fail until you list their hosts; WebFetch, MCP servers and Mode B stay open.
  • Protection follows the path: protected data moved out of a protected folder is not protected.

docs/threat-model.md lists every limit.

Documentation

Use phiward

Decide whether phiward is enough for your data

  • Why phiward: what exists already, and the gap it fills.
  • The two modes: what each one confines.
  • The threat model: what is protected, what is not, every known limit (a summary for reviewers comes first).

Change phiward

In a project that uses phiward, phiward init adds the rules for agents to that project's AGENTS.md (the text is in src/phiward/data/agents-section.md), and skills/phiward/SKILL.md is the skill.

Licence

AGPL-3.0-only. See LICENSE. Design credits are in NOTICE.

Metadata

Release files for phiward 0.2.0b1

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for phiward 0.2.0b1
File Size Uploaded Provenance Actions
phiward-0.2.0b1.tar.gz 670.7 kB With provenance Details for phiward-0.2.0b1.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for phiward 0.2.0b1
File Interpreter ABI Platform Actions
phiward-0.2.0b1-py3-none-any.whl Python 3 none any Details for phiward-0.2.0b1-py3-none-any.whl

Total release size: 1.0 MB

Release files / phiward-0.2.0b1.tar.gz

Download URL phiward-0.2.0b1.tar.gz
Size 670.7 kB
Tags Source
SHA-256 checksum
How to use checksums
043cf96e5b09f15fb28c09b1943f3d988c853ed950159cf3c87b0ca02432ec45
BLAKE2b-256 checksum
How to use checksums
351fa9749eb84c0b496c94c01be240101d438b8c360349e623246b0a2a9a2c6b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 18, 2026.

Transparency log

Release files / phiward-0.2.0b1-py3-none-any.whl

Download URL phiward-0.2.0b1-py3-none-any.whl
Size 338.3 kB
Tags Python 3
SHA-256 checksum
How to use checksums
18accb2a7ea80b8472791f6f13aa78d3a8aaeb306d93bbd7cbcc9e0de2b99310
BLAKE2b-256 checksum
How to use checksums
a9484a267a74f0c0cfd07d66f1352c17bc0f348a68966e7c7736cd83501a68f0
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 18, 2026.

Transparency log

Security for phiward 0.2.0b1

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for phiward-0.2.0b1.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 18, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release
Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page