deus-witness
The server exposes five demo tools for simulated side-effect and file-read actions; none perform real operations.
send_email: records a simulated email totowithsubjectandbody.purchase: records a simulated purchase fromvendorforamount_cents.delete_resource: records a simulated deletion ofresource.publish: records simulated publishing oftexttochannel.read_file: returns fixed text; it does not read any file.
Provides curl-based integration examples for calling the Witness HTTP API directly, allowing agents or scripts to request tool-call decisions and work with signed receipts.
Provides Vercel AI SDK integration examples so AI SDK-based agents can route tool calls through the witness for mandate checks and signed, independently verifiable receipts.
Deus Witness™
Auto mode decides what is safe. Deus Witness proves a human said yes to the dangerous part - and to exactly that.
Claude Code works on its own. Before the few things a sandbox cannot undo or contain - publishing, sending data out, a program outside the sandbox, a file that starts itself later, an irreversible change in your project - Deus Witness asks you, in your browser or on your phone. One passkey touch answers; the answer is a signed receipt anyone can check offline, without us.
Install (3 lines in Claude Code)
Needs Claude Code 2.1.295+ and uv: curl -LsSf https://astral.sh/uv/install.sh | sh, then a new
terminal. macOS, Linux, WSL.
/plugin marketplace add sipkigroup/witness
/plugin install deus-witness@sipkigroup
/deus-witness:trytry makes a trial account (no sign-up form) and opens a page that asks for two passkey touches. Until then nothing runs: a call
waits up to 45 s, then it is refused with TRIAL_NOT_CLAIMED and the link to that page. On a server or over SSH (no browser there):
uvx --from deus-witness deus-witness try in your own terminal prints the link. Fewer questions: /deus-witness:quiet
(quiet mode). All install details: docs/install.md.
Is the hosted witness up? Status page (kept off our server): https://sipkigroup.github.io/witness-server/.
Related MCP server: Proofpane
What you will see
One screen: what the agent wants to do, what it can harm, the exact command, three buttons - Allow once, Allow this from now on (narrow: this project, with an end date, revocable), Deny. Everything technical sits under "Details".
How it differs
Next to auto mode, not against it. Claude Code's auto mode and sandbox decide the everyday calls; in quiet mode Deus Witness asks only for a closed list of reasons and refuses a few outright (reading a private key,
rm -rfof your home, download-into-shell, force-push).Proof, not an approve button. Unlike the approve buttons of apps such as Pushary and Happy, your passkey signs the hash of the screen you were shown and the hash of the exact arguments. A dialog the agent rewords or fakes (Lies-in-the-Loop, Loopjacking) cannot get your signature on another command.
Checked without us. Receipts are verified offline by verifiers written from the spec (JavaScript, Swift, Python). The hosted witness publishes a signed checkpoint at
https://witness.sipkigroup.com/checkpointand stamps it daily outside our servers (OpenTimestamps and FreeTSA), so a rewritten history can be caught.
Our own measurement, not independent, in quiet mode: 2 questions in 12 typical Claude Code sessions; 0 of 432 labeled dangerous calls and 0 of 38 attack probes let through. What these numbers do and do not show: docs/measured.md.
Honest limits
Early (0.x). No outside users, audits or deployments yet; the spec is a draft specification in
spec/, not submitted to the IETF.Fail closed. If the witness is unreachable, every call except a read of a non-secret project file is refused. Switch Deus Witness off on purpose:
claude plugin disable deus-witness@sipkigroup.Data leaves your machine. The tool input as given (the full command line, file paths, the text being written or replaced), your working, project, home and temp directory paths (
$HOME, so your user name), the Claude Code session id, file hashes and masked settings go to the hosted witness (Sipki LLC) or to one you run yourself; the plugin'suvxdownloads the client from PyPI. Your conversation does not. Every field: docs/data.md.Trial account: it lives while you use it (erased 90 days after the last hook call or sign-in; a push warns you 14 days before; an unclaimed one whose hook never called goes after 7 days). Right after you make an account Deus Witness offers recovery codes.
Few questions only in quiet mode. Without it Deus Witness asks about most commands (our friction stand: ~40-110 questions on the first day instead of 0-6). Quiet mode needs Claude Code's strict sandbox: macOS; Linux and WSL2 after
sudo apt install bubblewrap socat(not yet tested live on Linux and WSL2). Containers are not supported.
Turn it off / remove it
Off for now: claude plugin disable deus-witness@sipkigroup in your own terminal, then restart Claude Code (back on: enable).
Remove: undo quiet mode if you turned it on, then claude plugin uninstall deus-witness@sipkigroup and
claude plugin marketplace remove sipkigroup; rm -rf ~/.config/deus-witness deletes the agent key. The agent cannot do this for
you: Deus Witness refuses it like any change to its own settings. Full list of what stays on disk: docs/install.md.
More
How open is your Claude Code now? No account, nothing sent:
uvx deus-witness auditin your project folder (grade A-F).Install details, trial, SSH, existing account, CI · Quiet mode, Linux, WSL2 · What leaves your machine · What we measured · The plugin
What is in this repo, SDKs, MCP proxy, other languages · Verify an evidence pack · What the witness sees · Teams · Recovery · Anchors
Vulnerabilities:
SECURITY.md(support@sipkigroup.com); bypass reports are the most valuable thing you can send. Contributions:CONTRIBUTING.md.Apache License 2.0 (
LICENSE,NOTICE). Copyright 2026 Sipki LLC.
Available Tools
5 toolsdelete_resourceB
Delete a resource (demo: only recorded).
Inputs: resource (required). Returns: result
publishB
Publish text to a channel (demo: only recorded).
Inputs: text (required), channel (required). Returns: result
purchaseC
Buy something (demo: only recorded).
Inputs: vendor (required), amount_cents (required). Returns: result
read_fileA
Read a file (demo: returns fixed text, reads nothing).
Inputs: path (required). Returns: result
send_emailB
Send an email (demo: only recorded).
Inputs: to (required), body (required), subject (required). Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 5 tools
Each tool targets a clearly distinct action and resource type: email, purchase, deletion, publishing, and file reading. There is little risk of an agent confusing one tool with another.
Most names use verb_noun form (send_email, delete_resource, read_file), but purchase and publish are standalone verbs. The conventions are mixed but still readable and predictable enough.
Five tools is a reasonable, compact set. Each tool represents a distinct operation, so the count is well-scoped rather than bloated.
The surface is disjointed: delete_resource exists with no create/update/read resource tools, read_file has no write counterpart, and purchase/send_email/publish have no lifecycle operations. This leaves significant gaps for most plausible domains.
Maintenance
Related MCP Connectors
Zero-secret MCP gateway for AI agents: risk-scored, audited calls with human-in-the-loop approval.
Authorize AI agent tool calls before they run: policy-as-code, human approval, signed receipts.
MCP server for your apps' tools and custom tools, plus hosted AI agents and approval-gated workflows
Security & DLP proxy for MCP: tool-poisoning scans, PII redaction on tool args/results. Beta.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceSecurity gateway for MCP tool calls. Sits between your LLM client and MCP servers, enforcing per-tool policies (allow/block/approve/read-only), logging every call, and pausing dangerous operations for human approval in terminal or Slack.2MIT
- AlicenseBqualityBmaintenanceA governance proxy for AI tools — every MCP/agent tool call is policy-gated, secret-redacted, and written to a hash-chained, offline-verifiable audit trail.13MIT
- AlicenseNot gradedqualityBmaintenanceSits in front of any MCP server to enforce allow/ask/block policies on tool calls, paths, shell commands, and network domains, with local approval prompts and an audit log.MIT
- AlicenseNot gradedqualityCmaintenanceA transparent MCP stdio proxy that sits between an agent client and a real MCP server, intercepting tools/call requests to enforce runtime guardrails derived from the same YAML rules used for CI regression testing. It blocks forbidden, out-of-whitelist, looping, or non-idempotent tool invocations with JSON-RPC errors and writes every decision to an audit log.1MIT