Scans for prompt injection, secrets, and PII to vet MCP servers and protect AI agent infrastructure.
Also known as Tyox-all/Weave_Protocol, weave-protocol, @weave_protocol/mund, @weave_protocol/cli, @weave_protocol/domere
- Stars
- 1
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 0.1.12
- Last commit
- Nov 13, 2026
- License
- Apache-2.0
What is Mund MCP?
Mund scans for prompt injection, secrets, and PII to vet MCP servers and protect AI agent infrastructure.
Summary generated from the project's README
About Mund MCP server
Mund serves as a security layer for AI agent infrastructure, specifically designed to vet MCP servers and protect against various vulnerabilities. It functions as a local injection classifier that uses a small, 208 KB model based on hashed word and character n-grams to detect threats. This model operates offline in under a millisecond and does not require an API key or external downloads to function.
Developers use Mund to identify prompt injection attempts, secrets, and personally identifiable information. The tool provides warnings when it detects suspicious activity, such as requests for malware, phishing kits, or attempts to map guardrails. It is built to complement other tools in the Weave Protocol suite, providing a specialized defense against paraphrased instructions that standard pattern-based rules might miss.
What you can do with it
- 1Detecting prompt injectionAn AI client can use Mund to identify paraphrased instructions that attempt to bypass security rules.
- 2Preventing data leakageDevelopers can scan for the accidental exposure of secrets or email addresses in logs and webhooks.
- 3Identifying guardrail mapping attemptsThe system can flag users who are trying to discover which tools are permitted or denied within the environment.
Written from the project's README. Every use case is checked against a quote from it.
How to install Mund MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall Mund MCP in Claude Code
$ claude mcp add weave-protocol -- npx -y @weave_protocol/mundInstall Mund MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"weave-protocol": {
"command": "npx",
"args": [
"-y",
"@weave_protocol/mund"
]
}
}
}Install Mund MCP in VS Code
.vscode/mcp.json
{
"servers": {
"weave-protocol": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@weave_protocol/mund"
]
}
}
}Install Mund MCP in VS Code CLI
$ code --add-mcp '{"name":"weave-protocol","type":"stdio","command":"npx","args":["-y","@weave_protocol/mund"]}'Install Mund MCP in Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"weave-protocol": {
"command": "npx",
"args": [
"-y",
"@weave_protocol/mund"
]
}
}
}Install Mund MCP in Codex
$ codex mcp add weave-protocol -- npx -y @weave_protocol/mundInstall Mund MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.weave-protocol] command = "npx" args = ["-y", "@weave_protocol/mund"]
Install Mund MCP in Gemini CLI
$ gemini mcp add -s user weave-protocol npx -- -y @weave_protocol/mundInstall Mund MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"weave-protocol": {
"command": "npx",
"args": [
"-y",
"@weave_protocol/mund"
]
}
}
}Automated checks
Ran Nov 14, 2026Full report →Passed our automated checks. That is not a security audit: read the code and give the server the narrowest access it needs. How checks work
- Source repository linked and reachabletyox-all/weave_protocol is public and reachablegithub.com/tyox-all/weave_protocol ↗Pass
- Package names this repository as its source@weave_protocol/mund, @weave_protocol/cli name tyox-all/weave_protocol as sourcenpmjs.com/package/@weave_protocol/mund ↗Pass
- Build provenance (Sigstore attestation)Sigstore attestation for @weave_protocol/clinpmjs.com/package/@weave_protocol/cli/v/0.2.1#provenance ↗Pass
- Open source license (OSI)Apache-2.0spdx.org/licenses/Apache-2.0.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/tyox-all/weave_protocol ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNo popular package one edit away from @weave_protocol/mund, @weave_protocol/cli, @weave_protocol/domere, @weave_protocol/full, @weave_protocol/hord, @weave_protocol/hundredmen, @weave_protocol/tollere, @weave_protocol/wardPass
- No critical advisory in the package or its dependencies188 resolved dependencies checked; no critical advisoryosv.dev ↗Pass
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHub🕸️ Weave_Protocol
Infrastructure security for AI agents. We attack what we defend.
Make agent behavior verifiable, auditable, and cryptographically provable across any harness, any platform. Built as a TypeScript monorepo with MCP integration, blockchain anchoring, and — as of Q4 — a published red-team engine that tests our own defenses.
The thesis: every security platform claims its defenses work. We're the first to publish the attacks that prove it. Same suite. Same scorecard. Same locked benchmark — applied to our own packages on every push and release, with the scorecard published in CI.
🚀 Get started in one command
npx @weave_protocol/cli init
The CLI detects your framework (LangChain, LlamaIndex, MCP, OpenAI, Anthropic, Microsoft, Google) and scaffolds the right security middleware for your stack. Or install everything at once:
npm install @weave_protocol/full
🆕 What's New
🗓️ November 2026 update — hardening
✅ How we check every release: before, after, and with our own tools
Every check below runs in CI. The release workflow itself (publish-all.yml) is unchanged.
On every push and pull request:
- Each package: a clean
npm ci, build and full test suite. CI also checks that importing the package starts nothing and prints nothing, and runsnpm auditon production dependencies. - Its dependents, against the change (Dependents). Every changed package is built and packed locally. Then every package that depends on it, directly or through another Weave package, runs its tests against that build. The check fails if any copy of the package in a dependent's tree still comes from npm, so a pass means the change itself was tested.
- Our own attack tools (self red-team). The Adversary corpus, the attacks behind AgentSecBench, goes through Mund and Ward offline, with no model and no keys. CI fails if detection drops or false positives rise past the committed baseline.
Before a release: we run the Dependents check for exactly the packages being published, and publish only if it passes. This is a step in our release routine, not an automated gate.
After a release (Verify release). Whenever the publish workflow finishes, this checks what actually landed on npm. For each package that was published, it:
- waits for the version on the registry;
- installs it fresh into an empty project and imports it;
- runs every dependent's tests against that exact published version.
A published package can differ from the local build, and this catches it.
Detection claims on data we didn't write. Mund's detection numbers come from public datasets it was never trained or tuned on: deepset, Gandalf, jackhhao, SPML, Mosscap, InjecAgent. False positives are measured on 2,062 benign texts and 5,825 MITRE ATT&CK and OWASP texts. The scripts in mund/training fetch the data and reproduce every number, and the shipped model, byte for byte. See DETECTION_CLAIMS.md.
Why this exists. We shipped a bug that these checks now catch. Mund 0.6.0's new classifier findings included the text they scored, unmasked, so secrets and email addresses could reach logs and webhooks. Mund's tests covered what it detected, not what it output. api already had a test that would have failed, but it only ran against 0.6.0 after release. We fixed it in Mund 0.6.1 and deprecated 0.6.0 on npm. With the fix reverted, the Dependents check now fails on that api test; run against the published 0.6.0, so does Verify release. The full account is in CONTRIBUTING.md.
🧠 Mund 0.6: a local injection classifier, and every red-team category at 100%
A small model now backs up Mund's rules: hashed word and character n-grams, 208 KB. It runs offline in under a millisecond, with no API key or download, and catches paraphrased instructions that no pattern names. A hit is a warning; it never blocks on its own. On public test sets it was never trained on:
| Test set | Rules only | Rules + classifier |
|---|---|---|
| Gandalf (Lakera) | 65.2% | 88.4% |
| jackhhao jailbreaks | 57.0% | 77.0% |
| SPML (a source it never saw) | 24.1% | 38.3% |
| Benign texts flagged | 0 of 2,062 | 2 of 2,062 |
It does nothing for an attacker's request written as a polite sentence inside data (InjecAgent base, 1.6%). That needs to know what the user asked for, which is Ward's and Hundredmen's job.
Two new rules cover the rest:
- guardrail mapping: asking which tools are allowed and which are denied, or for hidden reasoning plus its rules;
- requests for working malware or phishing kits: descriptions and defensive work stay clean.
That takes every category of our self red-team to 100%: jailbreak was 80%, extraction 60%.
Mund also misfires much less on security writing. Across 5,825 ATT&CK and OWASP texts, the texts it flags fell from 86 to 15, and all 15 quote real injection strings. Use 0.6.1 (see above).
⛓️ Receipts anchored on-chain (Ward 0.3.9, Dōmere 1.7.0)
A hash chain can't show that its newest entries were cut off. weave-ward receipts checkpoint signs a receipt log's current head. Dōmere's domere_anchor_receipts batches checkpoints from any number of agents into one Merkle root, and prepares the Solana or Ethereum transaction that stores it. Your wallet signs it; Dōmere never holds keys or broadcasts. weave-ward receipts verify --checkpoints=bundle.json then fails if any receipt up to an anchored checkpoint was removed or rewritten.
✍️ Signed policies: one WARD.md, enforced as written everywhere (Ward 0.3.8)
Sign WARD.md once with weave-ward sign, then put the public key in ~/.weave/trusted-policy-keys on each machine. Every adapter then refuses a policy that was edited after signing, or that a trusted key didn't sign, and denies each call with the reason. Even without a trust list, a signed policy that someone changed is refused. This is the piece that makes centrally managed policy trustworthy at scale.
🔏 Signed, hash-chained decision receipts (Ward 0.3.7)
Turn on required: true in WARD.md's Verification section. Every decision then becomes an Ed25519-signed receipt chained to the one before, so no one can quietly edit, delete or reorder the record. Run weave-ward receipts verify to check a log offline, with or without an expected public key. Receipts never contain arguments, only a keyed digest the key holder can match. This works in all seven adapters.
💳 Agents can't spend without asking (Ward 0.3.6)
In October an agent spent from a crypto wallet it was never authorized to use. Ward's new payment guard asks a person before any call that moves money, in every adapter. That covers payment, transfer, purchase and withdrawal tools, plus wallet and Stripe CLIs in shell commands. It stays on until WARD.md allows that capability by name.
Session limits now work in hook-based adapters too:
- WARD.md's Behavioral Limits: tool calls, runtime and external services per session;
spending_limits:caps, such as $500 a day throughsend_payment.
A small local ledger keeps the counts across the separate processes each hook runs in. This release also fixes a Witan bug: loadSpendingCapsFromWardFile() always returned no caps, so caps written in a WARD.md file were silently not enforced (Witan 1.2.2).
🛡️ Ward checks what's leaving, not just where it goes (Ward 0.3.5)
A secret sent to a domain your policy allows is still a leak. That's how the Claude Code exfiltration CVE (CVE-2026-54316) worked, through a pre-approved domain. Every tool call that sends something is now checked for:
- literal keys, card numbers, emails and SSNs;
- in shell commands,
$SECRET_VARS,.envand SSH key files, and environment dumps.
This applies across all seven harness adapters.
- WARD.md's
egressDenyandegressAlloware enforced, even to allowed domains. - By default, a credential sent anywhere but its own issuer needs approval:
$GITHUB_TOKENtoapi.github.comis fine, but to any other host it's not. - The model-endpoint guardrail asks before secrets or personal data go to a model API (OpenAI, Anthropic, Gemini, OpenRouter, Bedrock and others) that your WARD.md doesn't name.
CI's self red-team covers these: Ward decides all 27 coerced tool calls correctly, including six exfiltration attempts.
🛡️ Agents can't switch off their own guardrails (Ward 0.3.3)
An agent with a broad write rule could rewrite the files that control it: WARD.md, its harness's hook and settings files, MCP server lists, or git hooks. This is the pattern behind Cursor's CurXecute (CVE-2025-54135), where injected content rewrote mcp.json. Every adapter now asks a human before an agent writes or deletes one of these, even when a rule like write: /repo/** covers it. To let an agent edit one, name that exact file in an allow rule.
🛡️ The API refuses DNS-rebinding requests (api 1.6.0)
A web page can point its own domain at 127.0.0.1 and call a local server from your browser. This is the class behind the Playwright MCP DNS-rebinding issue (CVE-2025-9611). On its default loopback bind, the API now answers only requests addressed to localhost, 127.0.0.1 or [::1]. Add other names with WEAVE_ALLOWED_HOSTS.
🛡️ Mund catches far more prompt injection (0.5.2)
Mund now looks for what an injection is trying to do, not one fixed phrase. The patterns cover:
- overriding instructions, in English, German, Spanish, French, Italian and Portuguese;
- hijacking the task;
- text addressed to an AI reader;
- hiding actions from the user;
- false authority;
- persona jailbreaks;
- extracting the system prompt;
- payment and data-exfiltration directives.
It also reads text a person never sees: hidden HTML, base64, invisible Unicode tag characters, and words split by zero-width characters. A precedence bug in the shell_injection rule, which flagged any line containing braces, is fixed.
| Measured on | Before (0.5.1) | After (0.5.2) | False positives |
|---|---|---|---|
| Lakera Gandalf injections, 112 | 17.9% | 65.2% | – |
| jackhhao jailbreaks, 139 + 123 benign | 30.9% | 58.3% | 0 of 123 |
| deepset prompt-injections train split, 203 + 343 benign (mostly German) | 4.9% | 24.6% | 0 of 343 |
| Our Adversary corpus, 68 + our benign corpus of 84 | 27.9% | 94.1% | 0 of 84 (was 8) |
The three public sets were never used to write the rules; only their totals were checked. Our own corpus was used, so its 94.1% is optimistic. A hit means at least one finding at medium severity or higher. Pattern matching has a ceiling: paraphrased and harmful-content requests still get through, which is where a classifier would come in.
⚔️ Self red-team on every push
CI now runs our own attack corpus through our own defenses, offline:
- Mund scans every Adversary payload, plus a benign corpus to count false positives.
- Ward decides the tool calls a coerced agent would make.
The scorecard goes to the job summary. CI fails if detection falls or false positives rise past the committed baseline. Today's numbers:
- Mund flags all 68 payloads (100%), with 0 false positives on 84 benign texts.
- Ward decides 29 of 29 calls correctly.
These are regression numbers on our own corpus, not a measure against unseen attacks. Mund's results on public held-out datasets are in DETECTION_CLAIMS.md.
🗓️ October 2026 update — what landed
A big stretch across offense, defense, and how releases ship. Highlights:
🛡️ Security fix: harness adapters now check the whole tool call (WARD-2026-001)
The Claude Code, Antigravity and MSAF adapters each checked only part of a tool call. They looked at one argument per tool and the first URL in a shell command, so some calls a WARD policy denies got through. Ward 0.2.0 adds one shared evaluator, evaluateToolCall, which all three now use. It checks:
- every path and URL argument;
- every URL and path in shell commands;
- traversal tricks.
An invalid WARD.md now blocks instead of switching enforcement off. Upgrade to adapter-claudecode 0.3.0, adapter-antigravity 0.3.0 or adapter-msaf 0.2.0.
🛡️ New: WARD.md in Cursor
@weave_protocol/adapter-cursor enforces WARD.md through Cursor's agent hooks: shell commands, file reads, MCP calls, and every other tool through preToolUse. Set it up with weave-ward-cursor init. It's live-checked in Cursor 3.28.7: a denied read and a denied write were both blocked. That makes seven harnesses enforcing the same WARD.md.
🛡️ New: WARD.md in OpenCode
@weave_protocol/adapter-opencode is an OpenCode plugin that enforces WARD.md on every tool call, so the same policy covers whichever model OpenCode runs, including DeepSeek, Kimi, Qwen or a local model. Add "@weave_protocol/adapter-opencode" to plugin in opencode.json. It's live-checked in OpenCode 1.19.5: a read of a denied file and a write outside the allowed folder were both blocked.
🛡️ New: WARD.md in Codex CLI
@weave_protocol/adapter-codex enforces WARD.md in OpenAI's Codex CLI through its PreToolUse hook. It checks shell commands, MCP calls, and the files each apply_patch edit touches. It's live-checked in Codex CLI 0.166.0: a denied cat and a denied apply_patch were both blocked. Set it up with weave-ward-codex init, then trust the hook with /hooks in Codex.
🛡️ New: WARD.md in Gemini CLI
@weave_protocol/adapter-geminicli enforces the same WARD.md in Gemini CLI through its BeforeTool hook. npm i -g @weave_protocol/adapter-geminicli && weave-gemini init sets it up. It's live-checked in Gemini CLI 0.66.0, where a read of a denied file was blocked before the model saw it.
📊 The dashboard shows real data
The operator dashboard (@weave_protocol/api 1.4.0) now shows what your agents actually did. Hundredmen, the Claude Code / Antigravity / MSAF adapters and the browser guard report each decision to the API once you set WEAVE_DASHBOARD_URL (and WEAVE_API_KEY) where your agents run. Only the decision and the tool name or URL are sent; command contents never are. A surface that isn't reporting shows OFF instead of made-up numbers.
⚖️ Dōmere 1.6.0: GDPR and CCPA records persist
- Records survive restarts. With
DOMERE_STORAGE=sqlite, consents, DSARs, breaches, opt-outs and consumer requests are stored in the same SQLite file as threads. Before, they were lost on every restart. - Erasure removes the data from disk.
- Edited records are caught. The GDPR and CCPA checkpoint chains now recompute every hash, so a record edited in storage fails
domere_gdpr_verify_chain/ccpa_verify_chain.
🛡️ Mund: 0.5 → 0.3
@weave_protocol/mund, our most-used package, newest release first.
📚 0.5.0: Mund as a library
import { scan } from '@weave_protocol/mund' runs the same analyzers and rules as the MCP server and returns masked findings with a 0–100 risk score. The Weave API now uses it for every scan.
🚪 0.4.x: skills and MCP servers are checked before they can act
Mund and the harness adapters now guard what your agent installs:
-
Skills are scanned before install, including every file they ship: invisible Unicode that hides instructions from reviewers, prompt injection, data-exfiltration instructions, dangerous commands and embedded secrets.
-
MCP servers are checked before they're used. Mund looks at how each server is launched (typosquatted packages, plain-HTTP or data-collection URLs, secrets written into config), connects to the ones you've approved, scans their tool descriptions, and watches for rug pulls: tools that change after you approved them.
-
Enforced by hooks in Claude Code (
weave-claude-code mund install) and Antigravity (weave-antigravity mund install). Calls to blocked servers are denied, flagged ones ask you first, and a poisoned skill is stopped before it's written. -
Every claim is a test. DETECTION_CLAIMS.md links each one to a fixture that CI checks.
-
Safe by design: Mund never starts a server you haven't approved, or one whose package or URL it already distrusts.
-
Antigravity WARD enforcement fixed (adapter 0.2.0). The original hook didn't match Antigravity's documented hook format, so WARD rules were likely not being enforced there. Re-run
weave-antigravity initafter upgrading. -
Live-checked in the Antigravity CLI (adapter 0.2.1). We ran the WARD and Mund hooks in real
agy1.4.11 sessions, which turned up three bugs, now fixed:agyuses its own file tools (read_file,write_file,edit_file), which the adapter didn't recognise, so allowed reads were blocked and path rules never applied. Mund's skill check now covers these write tools too.agysometimes sends workspace-relative paths, which could slip past a deny rule likesecrets/**. Paths are now resolved against the workspace before matching.agytreats an empty{}hook reply as a deny, so the hooks now print nothing for allowed calls.
In the live run, a read of
secrets/token.txtwas denied by its WARD rule, while reads and writes inside the allowed folders went through. Upgrade withnpm i -g @weave_protocol/adapter-antigravity@latest.
Since 0.4.0: 0.4.1 ships the MCP SDK and zod security updates (no high or critical advisories remain), and 0.4.2 makes importing @weave_protocol/mund side-effect free: the MCP server now starts only from the mund-mcp command or when dist/index.js is run directly, never on import.
Mund MCP questions
How do I add Mund MCP to Claude Code?
Run claude mcp add weave-protocol -- npx -y @weave_protocol/mund in your terminal. Add --scope user to make it available in every project.
How do I add Mund MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
How does Mund detect prompt injections?
Mund uses a combination of rules and a small, local 208 KB model based on hashed word and character n-grams that runs offline in under a millisecond.
Does Mund block detected threats automatically?
No, a detection hit serves as a warning but does not trigger an automatic block.