mcp-audit

A drop-in security and observability proxy for MCP servers. mcp-audit sits between an MCP client and any upstream MCP server to produce signed audit trails, redact sensitive payloads, enforce allow/deny policies and per-tool rate limits, and expose a local read-only dashboard.

For a contributor-oriented map of the runtime, package boundaries, concurrency model, and design invariants, see ARCHITECTURE.md.

Why mcp-audit?

The MCP 2026 roadmap calls out enterprise needs around audit trails, gateway patterns, and operational visibility. mcp-audit fills that gap as a deployable sidecar or local wrapper: it sits between any MCP client and server, preserves protocol traffic, and records signed audit entries for tool calls, resource reads, prompt requests, and all other JSON-RPC methods.

+-------------+    JSON-RPC / MCP     +-----------+      JSON-RPC / MCP     +-------------+
| MCP client  | <-------------------> | mcp-audit | <---------------------> | MCP server  |
+-------------+                       +-----------+                         +-------------+
                                            |
                                            v
                                JSONL or SQLite audit log
                                            |
                                            v
                                   Read-only dashboard

Related MCP server: GitHub MCP Server Plus

What This Is / Is Not

mcp-audit is not a domain-specific MCP server. It is a transparent security and observability proxy that wraps any MCP server and audits the JSON-RPC traffic passing through it.

Directories may show the tools exposed by the upstream server, not tools implemented by mcp-audit itself.

Supported Transports

• stdio for local MCP clients such as Claude Desktop

• http for MCP servers exposed over HTTP

HTTP upstreams can use custom CA bundles, TLS server name overrides, and optional mTLS client certificates. Upstream retries are disabled by default and only apply to conservative, idempotent JSON-RPC methods when enabled; tools/call is not retried.

Use Cases

• Audit tool calls made by AI agents in regulated environments

• Detect unexpected or dangerous MCP tool usage

• Keep signed JSONL or SQLite logs for incident review

• Redact sensitive fields before storing requests and responses

• Block disallowed tools and apply per-tool rate limits without modifying the upstream MCP server

Demo

Install

For detailed platform-specific instructions and troubleshooting, see INSTALL.md.

Download the latest prebuilt binary and run:

# Linux/macOS: resolve latest, download, verify it starts
version=$(curl -fsSL https://api.github.com/repos/P4ST4S/mcp-audit/releases/latest \
  | grep '"tag_name"' | head -n1 | cut -d'"' -f4 | sed 's/^v//')
os=$(uname | tr '[:upper:]' '[:lower:]')
arch=$(uname -m); [ "$arch" = "x86_64" ] && arch=amd64 || arch=arm64
base="https://github.com/P4ST4S/mcp-audit/releases/download/v${version}"
archive="mcp-audit_${version}_${os}_${arch}.tar.gz"
curl -L -o "${archive}" "${base}/${archive}"
tar -xzf "${archive}"
./mcp-audit --version

Run with Docker:

docker run --rm ghcr.io/p4st4s/mcp-audit:latest --version

Install from source with Go:

go install github.com/P4ST4S/mcp-audit/cmd/mcp-audit@latest

To pin a specific release for reproducible installs, see INSTALL.md.

Quick Start

Run in stdio mode:

AUDIT_SECRET="$(openssl rand -hex 32)" \
mcp-audit --transport stdio --upstream "npx @modelcontextprotocol/server-filesystem /tmp"

On Windows PowerShell, generate the secret and set it as an environment variable:

$env:AUDIT_SECRET = -join ((1..32) | ForEach-Object { '{0:x2}' -f (Get-Random -Max 256) })
.\mcp-audit.exe --transport stdio --upstream "npx @modelcontextprotocol/server-filesystem C:\Temp"

Run in HTTP mode:

mcp-audit --transport http --upstream http://localhost:8080 --port 4422

Run with Docker Compose:

docker compose up --build

The dashboard is available at http://127.0.0.1:9090 by default. Prometheus metrics are available at http://localhost:9091/metrics by default.

Examples

• Cursor stdio configuration

• Continue stdio configuration

• VS Code stdio configuration

• Claude Desktop stdio configuration

• Zed stdio configuration

Configuration

mcp-audit loads config.yaml from the current directory by default. CLI flags override config values, and AUDIT_SECRET overrides audit.secret.

By default, mcp-audit strips hop-by-hop request headers and Authorization before forwarding HTTP requests to the upstream. To pass a bearer token to a trusted authenticated upstream, opt in explicitly:

proxy:
  forward_headers:
    - Authorization

Security note: forwarded headers, including secrets like bearer tokens, are transmitted verbatim to the upstream server. Only enable this if you control or trust the upstream MCP server. Authorization is the only sensitive header that can be opt-in forwarded because some MCP HTTP servers require it for upstream authentication. Cookie, Set-Cookie, and Proxy-Authorization are always rejected: they represent state destined for other components such as browser sessions or proxy chains and have no legitimate use in MCP request forwarding. If an existing deployment relied on implicit Authorization forwarding, add the config above.

JSONL rotation is disabled by default and supports size-based and UTC time-based triggers. Rotated archives use UTC timestamps such as audit.jsonl.20260610T214605Z; if multiple rotations happen in the same second, numeric suffixes are added. The archive timestamp reflects the wall-clock time of the rotation event, not the cutoff that was crossed. Time-based rotation is append-driven: mcp-audit does not start a background timer, so if no writes occur for several days, the active file is not rotated until the next append after the cutoff. Missed cutoffs are not caught up; the next append creates at most one archive.

audit:
  storage: jsonl
  rotation:
    max_size_bytes: 104857600
    interval: daily
    max_files: 10
    max_age_days: 30

max_age_days uses the rotation timestamp encoded in the archive filename. This means age since rotation, not the age of the oldest entry inside the archive. Age retention runs before max_files retention. Compression and SQLite archival are not part of this release.

CLI flags:

--transport    stdio | http
--upstream     upstream server command or URL
--port         proxy port for http mode
--upstream-timeout upstream HTTP request timeout in milliseconds
--config       path to config.yaml
--storage      jsonl | sqlite
--no-dashboard disable the web dashboard
--no-metrics   disable Prometheus metrics
--version      print version and exit
--log-level    debug | info | warn | error

Claude Desktop

Configure Claude Desktop to spawn mcp-audit instead of the upstream MCP server:

{
  "mcpServers": {
    "filesystem-audited": {
      "command": "mcp-audit",
      "args": [
        "--transport",
        "stdio",
        "--upstream",
        "npx @modelcontextprotocol/server-filesystem /tmp"
      ],
      "env": {
        "AUDIT_SECRET": "replace-with-a-long-random-secret"
      }
    }
  }
}

Dashboard

The dashboard shows recent entries, filters, expandable request/result JSON, top tools, calls today, and error rate. It refreshes every five seconds.

By default the dashboard listens only on 127.0.0.1:9090. To expose it on another interface, configure dashboard.bind_address explicitly and enable authentication or place it behind a trusted access proxy.

dashboard:
  enabled: true
  bind_address: 127.0.0.1
  port: 9090
  auth:
    token: "replace-with-a-long-random-token"

When dashboard.auth.token is configured, requests to /, /api/entries, and /api/stats must include:

Authorization: Bearer replace-with-a-long-random-token

Missing or invalid credentials return 401 Unauthorized with WWW-Authenticate: Bearer realm="mcp-audit-dashboard". Repeated failed authentication attempts from the same remote address are throttled with 429 Too Many Requests.

Dashboard JSON API responses include Cache-Control: no-store so intermediaries and browsers do not retain audit payloads.

Prometheus Metrics

mcp-audit exposes Prometheus metrics on a separate endpoint so platform teams can scrape operational data without exposing the dashboard.

scrape_configs:
  - job_name: mcp-audit
    static_configs:
      - targets: ["localhost:9091"]

Application metrics use the mcp_audit_ prefix and avoid unbounded labels. Tool-level labels can be disabled with metrics.tool_labels: false for stricter cardinality control. Policy decisions are exposed as mcp_audit_policy_decisions_total{action="allow|deny"}.

For a ready-made Prometheus + Grafana stack, see examples/docker-compose-observability.

Policy Engine

mcp-audit can enforce synchronous allow/deny rules before a tools/call reaches the upstream MCP server. Denied calls return a JSON-RPC error and are still written to the audit log.

policy:
  enabled: true
  default_action: allow
  rules:
    - action: deny
      client_id: claude-desktop
      server_id: filesystem
      tool_name: delete_file
      reason: "Destructive filesystem operations are blocked"

Rules are evaluated in order. Empty fields and * match any value, so default_action: deny can be used with explicit allow rules for stricter deployments.

OpenTelemetry

mcp-audit can export tools/call audit entries as OTLP/HTTP JSON spans to Jaeger, Tempo, Honeycomb, or any OTLP-compatible collector.

otel:
  enabled: true
  endpoint: "http://localhost:4318"
  service_name: "mcp-audit"
  headers:
    Authorization: "Bearer your-token"
  timeout_ms: 5000
  retry:
    max_retries: 3
    initial_interval_ms: 200
    max_interval_ms: 2000

The exporter uses current OpenTelemetry MCP and GenAI semantic conventions where possible, including mcp.method.name, jsonrpc.request.id, gen_ai.operation.name, gen_ai.tool.name, network.transport, network.protocol.name, rpc.response.status_code, and error.type. Project-specific attributes are kept link-oriented, such as mcp_audit.entry_id, mcp_audit.direction, mcp_audit.client_id, mcp_audit.server_id, mcp_audit.storage, and mcp_audit.signature.present.

Request params and tool results are not exported to spans by default. The signed JSONL or SQLite audit row remains the evidence artifact; OTLP provides correlation, latency, and operational visibility.

Exporter health is visible through Prometheus metrics under the mcp_audit_otel_ prefix, including export requests, span outcomes, dropped spans, queue depth, and queue capacity. Temporary OTLP failures are retried with bounded exponential backoff; Retry-After is honored for retryable responses up to otel.retry.max_interval_ms.

Audit Entries

Each stored entry includes a ULID, timestamp, direction, transport, JSON-RPC method, tool name when present, redacted params/result, JSON-RPC error when present, duration, client/server identifiers, and an optional HMAC-SHA256 signature.

Example JSONL entry:

{
  "id": "01HY8G6Y8S6W9K6ZD7VJ4Q8X4R",
  "timestamp": "2026-05-25T12:34:56Z",
  "direction": "client_to_server",
  "transport": "stdio",
  "method": "tools/call",
  "tool_name": "read_file",
  "params": {
    "name": "read_file",
    "arguments": {
      "path": "/tmp/example.txt",
      "token": "[REDACTED]"
    }
  },
  "duration_ms": 18,
  "client_id": "claude-desktop",
  "server_id": "filesystem",
  "signature": "hmac-sha256:..."
}

The signature covers:

id + timestamp + method + tool_name + raw_params

Roadmap

• SIEM-friendly exports

• OTLP compression and trace context propagation

Contributing

Keep changes small, run go build ./... and go vet ./..., and prefer standard library behavior over new dependencies. Stability guarantees are documented in STABILITY.md.

See CONTRIBUTING.md for setup, PR expectations, and project principles. See CHANGELOG.md for release history.

Community

• Discussions: questions, ideas, and design conversations

• Issues: bug reports and concrete feature requests

• Security: see SECURITY.md for the private vulnerability reporting process

License

Apache-2.0. See LICENSE.