epistemic-guard: Signed Tool Receipts & Verification
Verifies agent tool claims with Ed25519 receipts and dedupes syndicated RAG sources
- Pricing
- Open source
- Platforms
- CLI, Linux, macOS
- Runs
- Locally
- Version
- 0.4.2
- Listing updated
- today
Similar MCP servers
About epistemic-guard: Signed Tool Receipts & Verification
epistemic-guard: Signed Tool Receipts & Verification is an open-source MCP server for security. It runs locally on your machine and is started by your AI client.
The README says the package checks whether an assistant reply describes tool actions that never actually ran, such as a refund that failed or a file that was never written. It does this by comparing the reply against signed receipts that a proxy or decorator records for each tool call, including runtime errors and pre-execution rejections.
The MCP mode runs as a separate process that wraps any stdio MCP server, keeps the signing key out of the child environment, and appends signed entries to a Merkle JSONL log that can be verified from the command line or in CI. A Python mode offers decorators and a supervisor class for checking prose clause by clause, including streamed output. Memory writes can be gated so that only entries signed by user or tool keys are stored in a SQLite provenance store.
It is aimed at developers building MCP servers, tool-calling agents and retrieval pipelines who need post-execution checks rather than input filtering. Neural classifiers are English-only and need an optional model download.
Install
Run in a terminal
claude mcp add epistemic-guard -- uvx epistemic-guardAdd to claude_desktop_config.json
{
"mcpServers": {
"epistemic-guard": {
"command": "uvx",
"args": [
"epistemic-guard"
]
}
}
}Add to ~/.cursor/mcp.json
{
"mcpServers": {
"epistemic-guard": {
"command": "uvx",
"args": [
"epistemic-guard"
]
}
}
}Add to .vscode/mcp.json
{
"servers": {
"epistemic-guard": {
"type": "stdio",
"command": "uvx",
"args": [
"epistemic-guard"
]
}
}
}Add to ~/.codeium/windsurf/mcp_config.json
{
"mcpServers": {
"epistemic-guard": {
"command": "uvx",
"args": [
"epistemic-guard"
]
}
}
}Add to your client's MCP settings (most use this shape)
{
"mcpServers": {
"epistemic-guard": {
"command": "uvx",
"args": [
"epistemic-guard"
]
}
}
}Using several servers? Build one config file with the MCP config generator.
Tools this server provides (4)
Taken from the project documentation; the running server may differ.
- verify_narration
- Verifies an assistant reply against receipts.
- verify_receipt_log
- Verifies a signed receipt log.
- inspect_receipt
- Inspects a single receipt.
- corroborate_sources
- Collapses duplicate sources and counts independent corroboration.
What it does
- Signed tool receipts
- Records completions, runtime errors and pre-execution rejections as Ed25519-signed entries in an append-only Merkle log.
- Out-of-process MCP proxy
- Wraps any stdio MCP server in its own process so the signing key is not passed to the wrapped server.
- Post-execution claim checking
- Compares assistant replies clause by clause against receipts and flags actions that were not executed or failed.
- Memory write gating
- Stores only entries whose signatures verify against trusted public keys, with append-only SQLite triggers and a hash chain.
- RAG source deduplication
- Collapses near-duplicate syndicated snippets so repeated wire copy is not counted as independent corroboration.
Good for
- Catching an agent that claims it sent an email or deleted a file when the tool call failed or never ran
- Auditing signed tool-call logs from an MCP server in CI pipelines
- Preventing unverified assistant prose from being written into long-term vector or SQLite memory
- Counting independent sources before answering medical, legal or security questions from retrieved pages
Pricing
Open source
Free and MIT-licensed; no paid plan is mentioned.
Prices change. Check rollsolo.app before you buy.
epistemic-guard: Signed Tool Receipts & Verification FAQ
Does it replace pre-execution guardrails?
No. The README positions it as a complement to input and tool-call firewalls such as NeMo Guardrails, checking output and memory writes after execution.
How is it installed?
The README gives pip install epistemic-guard, with an optional epistemic-guard download-models step for the INT8 neural weights.
Does it make LLM API calls?
The README states it runs locally with zero LLM API calls.
What languages do the neural classifiers support?
The README says the neural classifiers are English-only.
Is epistemic-guard: Signed Tool Receipts & Verification free?
Yes. epistemic-guard: Signed Tool Receipts & Verification is open source, so you can use and self-host it for free.
How do I add epistemic-guard: Signed Tool Receipts & Verification to Claude Code?
Run this command in your terminal: claude mcp add epistemic-guard -- uvx epistemic-guard
What are the best alternatives to epistemic-guard: Signed Tool Receipts & Verification?
The closest alternatives to epistemic-guard: Signed Tool Receipts & Verification are Aga MCP Server, AgentProof, Protect MCP, Provared MCP executor and N0N4 Guardian.
Where this comes from
- Description: written from the project's README, the project's website and its MCP Registry entry.
- Install steps: from the official MCP Registry entry.