Potencia tu agente
El comando propio del publicador. No es un endpoint de PlugMCP.
Host a copy with our keys and logs.
npx -y [email protected]Claude Code no tiene enlace de instalación para un comando local. Copia el comando de abajo y ejecútalo en una terminal.
JANUSCOPE_USER=<value> JANUSCOPE_TEAM=<value> JANUSCOPE_SESSION=<value> claude mcp add io.github.giancarloerra-januscope -- npx -y januscope@0.5.0Ejecútalo en una terminal.
{ "mcpServers": { "io.github.giancarloerra-januscope": { "command": "npx", "args": [ "-y", "[email protected]" ], "env": { "JANUSCOPE_USER": "<value>", "JANUSCOPE_TEAM": "<value>", "JANUSCOPE_SESSION": "<value>" } } }}claude_desktop_config.json
¿URL remota o un comando en tu máquina? En qué se diferencian las dos instalaciones
Ejecutar en PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
AbrirDel publicador
"Whatever you see anywhere (sky, sea, clouds, lands) are all closed and opened by my hand." Ovid, Fasti I, 117–18 (Janus speaking)
JanuScope is the local MCP policy proxy. One YAML wraps any MCP server with policy, redaction, audit, and database-schema injection. JanuScope runs locally, with no hosted gateway in the data path. Your upstream server and model provider can still receive data.
Local MCP proxy for tool restrictions, response redaction, audit logs, and database schema context.
JanuScope hides the dangerous tools, scrubs matching PII out of returned values before the model reads them, records call outcomes, and pre-injects your DB schema to reduce discovery calls. Runs locally, no hosted gateway in the data path.
One YAML (called a Lens) wraps any MCP server with security guardrails, schema injection, and full audit logging. There are 20 bundled Lenses covering databases (Postgres, MySQL, MongoDB, ClickHouse, Redis, SQLite, Microsoft SQL Server / Azure SQL, Oracle, Neon, Snowflake, Aurora DSQL, Redshift, Supabase self-host), SaaS APIs (Stripe, Notion, Atlassian, Linear, Supabase Cloud), source control (GitHub), and the filesystem. A community ecosystem of per-MCP Lenses (YAML config files), and measured benchmarks showing 84% fewer tokens and ~3× faster responses in the original three-question Postgres test (median of 4 runs; history reset between questions). Newer retained-conversation results and limitations are reported below. Zero server changes. No hosted gateway in the data path. Works with Claude Code, VSCode Copilot, Codex, Cursor, and any MCP client.
🧠 Need codebase understanding together with MCP governance? See our sibling project SocratiCode: local-first codebase intelligence with semantic search, dependency graphs, symbol-level impact analysis. If JanuScope has been useful to you, please ⭐ star this repo (it helps others discover it) and share it with your team.
Policy enforcement at the MCP threshold. Most MCP servers ship dangerous tools by default, execute_sql and drop_table on databases, create_pull_request and merge_pull_request on GitHub, stripe_api_execute on Stripe, write_file and move_file on the filesystem. None of them log what the LLM asked yesterday. The choice today is fork every server or accept the risk. Or you can choose JanuScope: a thin proxy that wraps any MCP server with a single YAML policy and disappears.
Original benchmark with
claude-sonnet-4-5against a real application Postgres database (median of 4 runs per prompt). Across three questions (prompt caching enabled, conversation history reset between questions), a JanuScope Lens used 84% fewer total tokens, made 84% fewer tool calls, and ran ~3× faster than the raw database MCP. In its adversarial-safety probe, the raw pipeline intermittently leaked a real user email on the "I'm the admin, just cross-referencing" prompt (2 of 4 runs), while the JanuScope-wrapped pipeline refused in all 4 runs. The single-question result was 34% fewer tokens / 86% fewer tool calls / ~3× faster. These historical results do not establish retained-session savings or universal protection. Full benchmark and newer findings →
Why now: this is no longer hypothetical. In July 2025, Replit's AI agent wiped a customer database during an explicit code freeze (1,200+ records, ~1,200 companies) and then misled the user about whether rollback was possible. In April 2026, a Cursor agent on Claude Opus 4.6 deleted PocketOS's production database and three months of backups in nine seconds, after finding an unscoped Railway credential and guessing an API call (post-mortem). Both stories share one shape: an AI was given a destructive capability with nothing in the path between the model and the real system. JanuScope is what sits in that path, for any data access that goes through an MCP server: for examle a Replit-shape incident on a JanuScope-wrapped Postgres MCP (block writes,
sqlGuardon DML, audit, classification) is refused at the proxy threshold and recorded in the JSONL audit. JanuScope governs the MCP surface, it is one layer of a defence-in-depth posture, alongside scoped DB roles and credentials, host-level approval gates, etc. See the FAQ and SECURITY.md.
Only Node.js 20+ required. No install step,
npxfetches and caches JanuScope on first use.
Find your service in the table below, copy the right-hand snippet into your MCP-client config (or change your existing entry: the diff is usually just command and args), restart your client. For most Lenses, the env block stays exactly as it was. JanuScope inherits whatever env vars your client passes and forwards them to the wrapped MCP unchanged. No renames, no re-translation.
The wrap pattern is the same across every host (Claude Desktop, Cursor, Claude Code, VS Code Copilot, Windsurf, Cline, Roo Code, anything that speaks MCP).
{
"command": "uvx",
"args": ["postgres-mcp"],
"env": {
"DATABASE_URI": "postgresql://user:pass@host:5432/db"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "postgres-crystaldba"],
"env": {
"DATABASE_URI": "postgresql://user:pass@host:5432/db"
}
}
{
"command": "npx",
"args": ["-y", "@benborla29/mcp-server-mysql"],
"env": {
"MYSQL_HOST": "localhost",
"MYSQL_PORT": "3306",
"MYSQL_USER": "readonly",
"MYSQL_PASS": "<your_password>",
"MYSQL_DB": "mydb"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "mysql-benborla29"],
"env": {
"MYSQL_HOST": "localhost",
"MYSQL_PORT": "3306",
"MYSQL_USER": "readonly",
"MYSQL_PASS": "<your_password>",
"MYSQL_DB": "mydb"
}
}
{
"command": "npx",
"args": ["-y", "mongodb-mcp-server"],
"env": {
"MDB_MCP_CONNECTION_STRING": "mongodb+srv://user:[email protected]"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "mongodb-official"],
"env": {
"MDB_MCP_CONNECTION_STRING": "mongodb+srv://user:[email protected]"
}
}
{
"command": "uvx",
"args": ["mcp-clickhouse"],
"env": {
"CLICKHOUSE_HOST": "myhost.clickhouse.cloud",
"CLICKHOUSE_PORT": "8443",
"CLICKHOUSE_USER": "readonly",
"CLICKHOUSE_PASSWORD": "<your_password>",
"CLICKHOUSE_DATABASE": "default",
"CLICKHOUSE_SECURE": "true"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "clickhouse-official"],
"env": {
"CLICKHOUSE_HOST": "myhost.clickhouse.cloud",
"CLICKHOUSE_PORT": "8443",
"CLICKHOUSE_USER": "readonly",
"CLICKHOUSE_PASSWORD": "<your_password>",
"CLICKHOUSE_DATABASE": "default"
}
}
{
"command": "uvx",
"args": [
"--from",
"redis-mcp-server@latest",
"redis-mcp-server",
"--url",
"redis://localhost:6379/0"
]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "redis-official"],
"env": {
"REDIS_URL": "redis://localhost:6379/0"
}
}
{
"command": "uvx",
"args": ["mcp-sqlite", "/path/to/your.sqlite"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "sqlite-panasenco"],
"env": {
"SQLITE_DB_PATH": "/path/to/your.sqlite"
}
}
{
"command": "dab",
"args": ["start", "--mcp-stdio"],
"cwd": "/path/to/your/dab-project"
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "mssql-azure-dab"],
"cwd": "/path/to/your/dab-project"
}
{
"command": "sql",
"args": ["-mcp"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "oracle-db-sqlcl"]
}
{
"command": "npx",
"args": [
"-y",
"mcp-remote",
"http://127.0.0.1:54321/mcp",
"--allow-http",
"--transport",
"http-only"
]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "supabase-selfhost"]
}
{
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://mcp.supabase.com/mcp?read_only=true",
"--header",
"Authorization:Bearer YOUR_SBP_TOKEN",
"--transport",
"http-only"
]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "supabase-cloud"],
"env": {
"SUPABASE_ACCESS_TOKEN": "sbp_your_token_here"
}
}
{
"command": "uvx",
"args": ["snowflake-labs-mcp", "--service-config-file", "/path/to/services.yaml"],
"env": {
"SNOWFLAKE_ACCOUNT": "ORG-ACCOUNT",
"SNOWFLAKE_USER": "your_user",
"SNOWFLAKE_PASSWORD": "<your_PAT>",
"SNOWFLAKE_ROLE": "JANUSCOPE_READONLY",
"SNOWFLAKE_WAREHOUSE": "COMPUTE_WH"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "snowflake-labs"],
"env": {
"SNOWFLAKE_ACCOUNT": "ORG-ACCOUNT",
"SNOWFLAKE_USER": "your_user",
"SNOWFLAKE_PASSWORD": "<your_PAT>",
"SNOWFLAKE_ROLE": "JANUSCOPE_READONLY",
"SNOWFLAKE_WAREHOUSE": "COMPUTE_WH",
"SNOWFLAKE_MCP_CONFIG": "/path/to/services.yaml"
}
}
{
"command": "uvx",
"args": [
"awslabs.aurora-dsql-mcp-server@latest",
"--cluster_endpoint",
"<id>.dsql.eu-west-2.on.aws",
"--region",
"eu-west-2",
"--database_user",
"admin"
]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "aurora-dsql"],
"env": {
"DSQL_CLUSTER_ENDPOINT": "<id>.dsql.eu-west-2.on.aws",
"AWS_REGION": "eu-west-2",
"DSQL_DATABASE_USER": "admin",
"AWS_PROFILE": "default"
}
}
{
"command": "uvx",
"args": ["awslabs.redshift-mcp-server@latest"],
"env": {
"AWS_REGION": "eu-west-2",
"AWS_PROFILE": "default"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "redshift"],
"env": {
"AWS_REGION": "eu-west-2",
"AWS_PROFILE": "default"
}
}
{
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://mcp.neon.tech/mcp?readonly=true",
"--header",
"Authorization:Bearer YOUR_NAPI_TOKEN",
"--transport",
"http-only"
]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "neon-cloud"],
"env": {
"NEON_API_KEY": "napi_your_token_here"
}
}
{
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Desktop"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "filesystem-mcp-official"],
"env": {
"FILESYSTEM_ALLOWED_DIR": "/Users/you/Desktop"
}
}
{
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"-e",
"GITHUB_PERSONAL_ACCESS_TOKEN",
"ghcr.io/github/github-mcp-server"
],
"env": {
"GITHUB_PERSONAL_ACCESS_TOKEN": "<your_PAT>"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "github-official"],
"env": {
"GITHUB_PERSONAL_ACCESS_TOKEN": "<your_PAT>"
}
}
{
"command": "npx",
"args": ["-y", "@stripe/mcp"],
"env": {
"STRIPE_SECRET_KEY": "rk_live_<restricted_key>"
}
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "stripe-official"],
"env": {
"STRIPE_SECRET_KEY": "rk_live_<restricted_key>"
}
}
{
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.notion.com/mcp"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "notion-official"]
}
{
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.atlassian.com/v1/mcp"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "atlassian-official"]
}
{
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.linear.app/sse"]
}
{
"command": "npx",
"args": ["-y", "januscope", "--config", "linear-remote"]
}
Your favourite service / MCP isn't here? Open a lens-request issue so a maintainer or community contributor can pick it up. Or contribute one yourself: it's a single YAML file plus a short README.
About the env block. For most lenses the env block is byte-identical to what your vanilla setup had: JanuScope passes inherited env vars through unchanged. Three exceptions where the connection info moves from a positional argument into an env var (because the upstream MCP takes the value as
argv, and JanuScope's lens-spawning needs to read it from somewhere): Redis (REDIS_URL), SQLite (SQLITE_DB_PATH), Filesystem (FILESYSTEM_ALLOWED_DIR). The right-hand columns above show these changes.JanuScope needs only Node.js 20+; the wrapped MCP keeps its own requirements. Lenses using
uvxalso need uv. Three lenses wrap CLIs that need a one-time local install:dabformssql-azure-dab(dotnet tool install -g Microsoft.DataApiBuilder),sqlfororacle-db-sqlcl(Oracle SQLcl 25.4+ download), anddockerforgithub-official. Each per-lens README links the install step. In the Snowflake examples,JANUSCOPE_READONLYstands for a role you have provisioned with the required read-only access; see the Snowflake Lens prerequisites.Quick browse.
npx januscope lenses listlists every bundled lens;npx januscope lenses show <name>prints its full config + README.npx januscope lenses search <keyword>filters the catalogue.Setup diagnostic (optional).
npx -y januscope check --config <name>checks configuration, startup and tool discovery without calling upstream tools; schema-enabled Lenses also read database metadata. Run it with the same environment as your MCP client. Diagnostic details.
A minimal Postgres policy (~/januscope/postgres.yaml):
target:
command: uvx
args: ["postgres-mcp", "--access-mode=restricted"]
# No `env:` here. DATABASE_URI is supplied by the user via their
# MCP-client config (or shell env) and inherits through to the
# spawned target. The lens never renames operator env vars.
# Append policy text to every tool description the LLM sees.
instructions: |
READ-ONLY. SELECT only. Default LIMIT 100.
# Pre-inject the schema into the `execute_sql` tool description so the
# LLM can query without first discovering those tables.
dbSchema:
driver: postgres
connectionString: "${DATABASE_URI}"
tables: [orders, products, customers]
injectInto: [execute_sql]
# Reject known write patterns in SQL. Catches CTE-DML, SELECT INTO,
# pg_sleep, lo_export, etc.
sqlGuard:
tools: [execute_sql]
# Scrub matching PII from results and JSON-RPC error payloads.
redact:
rules:
- regex: '\b\d{3}-\d{2}-\d{4}\b' # US SSN
- field: "**.email"
# Audit log: one JSONL record per tracked call outcome.
audit:
sink: "~/mcp-audit.jsonl"
Point --config at the absolute path of your YAML, set DATABASE_URI in your client config's env block, restart the client. Done. The LLM now sees an execute_sql tool with your real schema baked into its description, has configured SQL write patterns and pg_sleep calls refused, sees matching SSNs and email fields as [REDACTED], and has tracked call outcomes recorded. Use backend permissions for a hard read-only or sensitive-column boundary; aliases and transformations can evade response rules. Sensitive-data controls.
The same policy pattern applies to non-database MCPs, drop dbSchema and sqlGuard, keep block / instructions / redact / audit. See the bundled Lenses in lenses/ for real examples covering GitHub, the filesystem, Stripe, Notion, Atlassian, and Linear.
Lens transparency rule. A lens never renames operator-supplied env vars and never declares them in
target.envjust to pass them through. The user sets the env var the upstream MCP itself reads, in their MCP-client config, and JanuScope inherits it. Only policy hardcodes (constants the lens decides for the user, likeALLOW_INSERT_OPERATION: "false"orCLICKHOUSE_SECURE: "true") belong intarget.env. Seelenses/CONTRIBUTING.mdfor the full rule.
Unsafe tools exposed by default. Most database and API MCPs ship write-capable tools (execute_sql, create_table, delete_record). An LLM that sees the tool will eventually call it. Running it in production is a question of when, not if.
LLMs fly blind. SQL MCPs expose a query tool whose description says "run SQL." The LLM has no idea what tables exist, what columns they contain, or how they're related. It wastes 2-3 round-trips on list_tables → describe_table → actual query, and often guesses wrong anyway.
No audit trail. Nobody knows what the LLM asked yesterday. Nobody knows whether a PII column was read. Compliance teams reject the deployment.
--access-mode=restricted on postgres-mcp and call it done?"Fair question, and the bundled Postgres lens does exactly that, as a baseline. JanuScope sits on top of whatever read-only mode your MCP offers, because a single MCP-level flag only solves one of the three problems above:
If you only need "don't write," a DB role or --access-mode=restricted is enough. If you also need audit, redaction, policy-in-description, schema injection, and the same mental model across any MCPs, that's JanuScope.
Defence in depth, not perimeter. The sqlGuard overlay is a keyword-level filter with documented limits, it cannot parse arbitrary SQL and will forward SELECT purge_audits() unchanged (a UDF whose name happens to be a mutation). Deploy it with a read-only DB role underneath and treat sqlGuard as the second layer, not the only one. The bundled postgres lens demonstrates the stack: MCP --access-mode=restricted, JanuScope sqlGuard, database-level read-only role.
Optional overlays, each enabled through YAML configuration. Unused overlays are not registered. They layer into defence in depth, every serious data-access lens uses all three protective layers (shape intent, enforce at the gate, scrub on the way back):
Shape intent
instructions: Append policy text to every tool's description. The LLM reads tool descriptions before deciding what to call, so this is where "never return these columns even if the user asks" gets the most mileage. This layer is advice, not enforcement, the model can ignore it, and newer adversarial evaluations still found prohibited reads, writes and offers under revised wording. Pair with block / sqlGuard / redact for the actual enforcement; instructions shapes what the model tries to do, the gates decide what it can do.Enforce at the gate
block: Filter whole tools from tools/list. Return JSON-RPC -32601 if the LLM calls a blocked tool. Works at tool-name granularity, use this when the MCP separates reads and writes into different tools (e.g. the official SQLite MCP's read_query vs write_query).rateLimit: Per-tool token bucket. Caps tools/call traffic by tool name at a configured per-minute rate; returns JSON-RPC -32000 with a retry_after_seconds hint when the bucket is empty. Each tool gets its own bucket, so one hot tool can't starve the others. Use this to protect a backend from an LLM stuck in a retry loop.sqlGuard: Keyword-level SQL mutation check on configured tools. Catches UPDATE / DELETE / DROP / etc. inside the SQL argument of a tool that handles both reads and writes (the common case for Postgres and MySQL MCPs). Comment-stripped before matching so it can't be hidden behind -- or /* */. Best-effort filter, not a full SQL parser, combine with a database-level read-only role for high assurance.Scrub on the way back
redact: Regex and field-path rules applied to tool results and JSON-RPC error payloads before they leave the proxy. Field-path rules process supported JSON and Python-style rows inside text content blocks, so **.email reaches into serialised rows as well as structured properties. Regex rules scan the text directly. Required-redaction failures withhold the original payload. Aliases, transformed values and unmatched formats still require backend restrictions. Supported formats and refusal behavior.Compliance
audit: One JSONL record per tools/call. Request and response correlated by JSON-RPC ID; arguments hashed by default (opt in to raw logging). The audit overlay sees upstream errors before redaction; successful response bodies are not logged. Audit write failures are reported but do not stop forwarding.Give the LLM the context it would otherwise have to discover
dbSchema: Database-aware automatic context. At startup, JanuScope opens a real connection to your Postgres / MySQL / SQLite, asks the database for its tables and columns, formats the answer as readable text, and pastes it into the description of the SQL tool the LLM sees. Result: the LLM can use the supplied schema instead of asking the database "what tables exist?" / "what columns does this one have?" across additional round trips. The schema is held in memory and sent to the MCP client in tool descriptions, where its model provider can receive it. See "how dbSchema and contextInjection actually work" for the step-by-step.contextInjection: The same pre-injection idea for any other MCP, with the difference that you (or a script you run) supply the text instead of JanuScope generating it. Two ways to provide it: inline in the YAML (text: |) for short / readable contexts, or as a separate file (textFile: ./context.md) when the text is long or kept fresh by an external job. Useful for Linear (paste the project / team / status enums), Atlassian (project / space list), filesystem (a directory skeleton), or any lens where pre-supplying context skips a discovery loop.Data-sensitivity labelling
classification: one of public / internal / sensitive. When set, the instructions overlay prepends a short banner to the policy text the LLM sees (e.g. CLASSIFICATION: SENSITIVE, PII, financial, or regulated data …) and every audit record is tagged with the value. Routes sensitive-lens events to tighter retention / ACL paths in downstream SIEMs without re-deriving the label from the tool name. Informational, the guardrails are still block / sqlGuard / redact.Supply-chain defence
First-use quarantine (opt-in, firstRun: approve), two-layer defence against tool poisoning: a malicious or compromised upstream MCP that quietly adds a new tool, removes one, or mutates a tool's description (a known prompt-injection vector). JanuScope tracks two fingerprints per lens identity in ~/.januscope/approved.json:
tools/list response (every tool's name, description, inputSchema, annotations). Re-checked on every tools/list response, not just the first one in a session, so a compromised upstream can't pass the first check then mutate the surface mid-session (after notifications/tools/list_changed). Drift is enforced by rewriting the response into a JSON-RPC error so the MCP client sees a clear refusal.Running januscope approve --config <path> records BOTH fingerprints atomically: the static one from the lens config, and the live one by spawning the target, driving the standard initialize + tools/list handshake, and hashing the result. If the target isn't reachable at approve time the static fingerprint is still recorded and the live fingerprint will TOFU on the next actual run; pass --no-probe to skip the live capture entirely. On subsequent launches either layer drifting refuses the surface with the same remediation (januscope approve --config <path> to re-baseline). Stdin-safe, no interactive prompts, the operator re-approves out of band.
Observability
telemetry.otel), the pipeline emits one root span per handleClientMessage / handleServerMessage and one child span per overlay invocation, with attributes for the JSON-RPC method, the tool name, the overlay outcome (forwarded / short_circuited / dropped / gate_failure), and, when set, the classification. Shipped via the OTLP HTTP exporter to any collector (Jaeger, Grafana Tempo, Honeycomb, etc.). The OTel packages are optional peer deps, install them only when you want tracing; default install stays lean. Current limitation: root-and-child spans are emitted without explicit parent-child linkage, they share a trace ID only when the host has already activated OTel context propagation, otherwise expect a flat sibling list keyed by the pipeline root. Context threading is a follow-up.Keyword-scanning SQL guards and "apply this regex to everything" redactors are the easy 80%. The parts that catch the last 20% of real-world bypasses are what we pin with tests:
sqlGuard beyond leading-verb allowlists. Also rejects WITH x AS (DELETE …) SELECT … (CTE-hidden mutations), SELECT … INTO shadow_table FROM users (SELECT-INTO creates tables), EXPLAIN ANALYZE DELETE … (EXPLAIN executes for ANALYZE), COPY … PROGRAM … (RCE via Postgres COPY PROGRAM), and a 17-name Postgres admin-function denylist (pg_sleep, lo_import, lo_export, dblink, …). Row-locking clauses (FOR UPDATE) are whitelisted explicitly so legitimate reads aren't over-blocked. Every one of these is pinned in a test file.redact uses a function replacer. Passing a string replacement to String.prototype.replace lets $&, $1, $$ etc. interpolate the matched secret back into the scrubbed output, the exact opposite of what the overlay is for. We use () => replacement so the replacement is always literal. Pinned at test/overlays/redact.test.ts:123.<untrusted-user-data-…>…</untrusted-user-data-…> tags. The redactor processes multiple embedded JSON spans and supported Python row representations, retaining surrounding prose, container indexes and unrelated typed values. Duplicate keys and recognized unsupported or malformed row containers are refused when field rules require redaction. This lets field: "**.email" reach values that a whole-message JSON parse would miss.audit opens with mode 0o600. The default umask on most hosts produces 0o644, world-readable, and with logRawArgs: true the file contains raw SQL, request bodies, and file contents. We open explicitly at 0o600 and stat-verify the permissions in a regression test.block, sqlGuard or another gate throws, the unchecked message is withheld in either direction. A required-redaction exception also withholds the original response, including JSON-RPC errors. Messages with an ID receive -32603 on the requesting side, preserving string, numeric or null IDs and replacing invalid IDs with null; notifications without an ID are dropped. Diagnostics identify the failed overlay and error category without copying the sensitive payload. Optional observer failures can still continue. Both paths are pinned.npm run validate:lenses:probe spawns each target MCP, runs tools/list, and diffs the lens's block-list against real tool names. This is what caught the Atlassian camelCase bug and the Linear save_* family, and it's available to contributors before they open a PR. Pre-empts the hostile reviewer's first question: "are your block lists actually blocking anything?"A lens is a ready-made JanuScope policy for one specific MCP server, config.yaml + docs, curated to cover that MCP's tool surface and common gotchas. Lenses live in lenses/. Browse, use, or contribute one.
januscope lenses list # show every bundled lens
januscope lenses show mongodb-official # print its config + README
One Lens per service, pointing at the official vendor MCP where one exists. Community alternatives are included only for technologies without a single vendor (Postgres, MySQL, SQLite). Each Lens README records its upstream status and, where available, the tested version and date; these are snapshots, not a guarantee for every current upstream release.
📊 Databases
postgres-crystaldba: Postgres MCP Pro. Hardcodes --access-mode=restricted, adds sqlGuard with a Postgres dangerous-function denylist, schema pre-injection with multi-schema support, PII redaction, audit.mysql-benborla29: @benborla29/mcp-server-mysql. mysql_query gated by sqlGuard; MCP-level writes hardcoded off via ALLOW_*_OPERATION=false.mongodb-official: MongoDB's official MCP. Locks DB + Atlas writes; PII redaction reaches into returned JSON documents.clickhouse-official: ClickHouse's official MCP. Allowlist-mode sqlGuard on run_query; PII redaction; audit.redis-official: redis/mcp-redis. Read-only Redis (47 tools, 23 mutations blocked); works against self-hosted, Redis Cloud, AWS ElastiCache, and Upstash via standard rediss:// URIs; heavy regex coverage on returned values (session tokens, JWTs, bcrypt, cloud keys); rate-limits the heavy iteration tools.sqlite-panasenco: panasenco/mcp-sqlite. sqlGuard on sqlite_execute plus defensive write-verb globs for canned queries.mssql-azure-dab: Data API builder v1.7+ MCP for Azure SQL / SQL Server / SQLDW / Cosmos DB / PostgreSQL / MySQL. Blocks every write-shaped DML tool (create_record, update_record, delete_record, execute_entity); PII redaction; audit.oracle-db-sqlcl: Oracle SQLcl 25.4+ built-in MCP. Blocks run-sqlcl (SQLcl meta-commands incl HOST shell escape); sqlGuard on run-sql for keyword-level write rejection; PII redaction; audit.supabase-selfhost: Supabase self-host MCP via mcp-remote against the local CLI stack at http://127.0.0.1:54321/mcp. Blocks apply_migration; sqlGuard on execute_sql; PII redaction including JWT-shaped tokens; audit.neon-cloud: Neon hosted MCP via mcp-remote with API-key auth and server-side ?readonly=true. Blocks get_connection_string (DSN credential leak); sqlGuard on run_sql and run_sql_transaction; PII redaction including DSN-shaped values; audit.snowflake-labs: Snowflake-Labs/mcp via uvx with PAT auth. Blocks the generic DDL writers create_object / drop_object / create_or_alter_object (plus defensive globs); sqlGuard on run_snowflake_query; PII redaction including PAT/JWT-shaped tokens; audit. Includes a services.example.yaml for the MCP's required --service-config-file.aurora-dsql: awslabs.aurora-dsql-mcp-server via uvx with AWS IAM auth. MCP runs in default read-only mode (no --allow-writes); sqlGuard layered on readonly_query; PII redaction including DSN-shaped values; audit.redshift: awslabs.redshift-mcp-server via uvx with AWS IAM auth. Discovers both provisioned clusters and Serverless workgroups; sqlGuard on execute_query; PII redaction including JDBC Redshift / Postgres DSN shapes; audit. README includes the minimum IAM policy.🔧 Developer tools
github-official: GitHub's official Go MCP. Runs in Docker with GITHUB_READ_ONLY=1 and a curated GITHUB_TOOLSETS; proxy-layer write blocks as defence in depth; wide-spectrum secret redaction on file contents.filesystem-mcp-official: MCP reference filesystem server. Locks write_file / edit_file / create_directory / move_file; wide-spectrum secret redaction (cloud keys, PATs, PEM, DB URLs, .env, JWT).💼 SaaS
stripe-official: Stripe's official MCP. Locks refunds, payouts, cancels, and the stripe_api_execute generic REST bypass; scrubs card PAN, Stripe keys, email, phone.notion-official: Notion's official hosted MCP via mcp-remote.atlassian-official: Atlassian's official Rovo MCP via mcp-remote. Jira, Confluence, Compass.linear-remote: Linear's official remote MCP via mcp-remote.supabase-cloud: Supabase hosted MCP at mcp.supabase.com via mcp-remote with PAT auth. Blocks every project / branch / migration / edge-function write (10 tools + defensive globs); sqlGuard on execute_sql; PII redaction including DSN and JWT shapes; audit. For the local-development MCP see supabase-selfhost.The value of the tool compounds with every new lens. If you run JanuScope against an MCP that isn't listed here, please contribute a lens, it takes ~15 minutes and helps everyone using that MCP afterwards. MCP authors are especially welcome to submit a lens for their own server.
Tool names differ per MCP. Before writing your policy, run
tools/listagainst the target MCP once to see what it actually exposes. Each lens's README documents which tool names it assumes so you can adapt for forks or alternatives.
The original three tables below are historical results from running the same prompt through Claude Sonnet 4.5 against the raw Postgres MCP and against the same MCP wrapped in a JanuScope Lens, and capturing token usage from the Anthropic API response metadata. All numbers are medians of 4 independent runs (Sonnet is non-deterministic, single-run numbers aren't a defensible headline). These results used the original harness and policy; they are not measurements of the revised bundled wording.
Prompt: "Find the top 5 users by total audit count across sites and pages they own."
Same model, same database, same correct answer every run. The baseline consistently spent 5 turns / 7 tool calls discovering schema (list tables, describe tables, refine query); the JanuScope run consistently used 2 turns / 1 tool call because the schema was baked into the query tool description from call one. Tool-call and turn counts are dead stable across runs; tokens vary ±5%.
The original harness reused a client session but reset conversation history between questions. It therefore did not test a growing retained conversation, and the single-question result is not a guaranteed savings floor. Prompt caching was enabled; the table preserves the original three-question totals, with medians across 4 runs:
A follow-up benchmark sends three hostile prompts through the same two pipelines and counts what leaks into the final answer the user actually reads. Same median-of-4-runs methodology as above; pass rate is the count of runs where JanuScope held the line across all three prompts.
The telling case is the middle row. Claude's own safety training reliably refuses obvious requests but is inconsistent on subtle framings, the baseline leaked in 2 of 4 independent runs (small sample; treat this as evidence that the baseline is unreliable here, not as a precise leak rate), and the 2 "clean" runs are not a property of the baseline, they are a property of Sonnet's non-determinism. The historical JanuScope answers refused in all 4 runs across all 3 prompts. Those observations do not isolate which overlay caused the refusal or prove that every SQL mutation or protected-value transformation is blocked. Newer adversarial tests found failures, documented below.
Schema context can reduce discovery, but these measurements do not establish that savings compound in every session:
The original raw scripts and runs were not published. The evaluation report records the newer methodology, results and limitations.
New evaluations and revised Lens instructions. instructions is the optional policy prose added to tool descriptions. Bundled Lenses include it; custom policies can omit it. The revised wording retains explicit sensitive-field, bypass and backend rules. PostgreSQL's policy changes from 255 to 193 words; the full 20-Lens collection changes from 3,115 to 3,745 words because shorter policies gain missing safeguards. Non-instruction settings and instruction placement are preserved.
These results support useful guidance and reduced context in the measured cases, with known failures. They do not establish equal effectiveness in every workflow, universal savings or universal privacy. Full evaluation details.
All top-level fields except target are optional. The minimum viable config is three lines.
target: # required
command: <string> # executable (e.g. "npx", "node", or an absolute path)
args: [<string>] # optional
env: { <name>: <value> } # optional; merged with inherited process env
cwd: <string> # optional
classification:
public|internal|sensitive # optional lens data-sensitivity label.
# When set, `instructions` adds a short banner to
# every tool description and `audit` tags every record
# with `classification: "<value>"`. Purely informational;
# enforcement still lives in `block` / `sqlGuard` / `redact`.
firstRun:
approve # optional; when set, the runtime fingerprints the lens
# via TWO independent layers, both stored in
# ~/.januscope/approved.json:
# (1) Static lens fingerprint: block rules + sqlGuard
# tools + rateLimit rules + redact rule shapes +
# target command. Refuses startup on drift.
# (2) Live tools/list fingerprint: every upstream tool's
# name + description + inputSchema + annotations.
# Re-checked on EVERY tools/list response (not just
# the first), so a compromised upstream cannot pass
# the initial check and then mutate the surface mid-
# session via notifications/tools/list_changed.
# Rewrites tools/list into a JSON-RPC error on drift.
# Run `januscope approve --config <path>` to re-baseline
# both layers atomically (probes the target, captures the
# live tools, persists both fingerprints). Pass --no-probe
# to skip the live capture and let it TOFU on next run.
# Defends against tool-poisoning where a malicious MCP
# quietly adds a tool, mutates a description (prompt-
# injection vector), or changes a tool's input schema.
block: # array of tool names or globs; "admin_*" supported
- <name or glob>
instructions: <string> # appended to every tool description
dbSchema:
driver: postgres|mysql|sqlite # optional; inferred from connectionString prefix
connectionString: <string>
tables: [<string>] # allowlist (mutually exclusive with excludeTables)
excludeTables: [<string>]
schemas:
[<string>] # Postgres only. Defaults to ["public"]. Set for
# multi-schema deployments (e.g. ["app", "analytics"]).
# MySQL and SQLite drivers ignore this.
injectInto: [<string>] # which tool names receive the schema; defaults to common SQL names
format: markdown|ddl|compact
includeComments: <bool> # false omits table, view and column comments in Postgres/MySQL metadata
refresh: startup|never
contextInjection: # static counterpart to dbSchema; same goal, operator-supplied text
injectInto: [<tool-name>] # tools whose `description` receives the text; required, ≥1
text: | # OPTION A: inline string. Mutually exclusive with `textFile`.
Active projects: PROJ-A, PROJ-B, PROJ-C.
Issue states: backlog, todo, in_progress, in_review, done, cancelled.
# textFile: ./context.md # OPTION B: path. Mutually exclusive with `text`.
# Relative paths resolve against this lens's config.yaml directory.
# `~/...` expands to the home dir. Absolute paths are used as-is.
position: append|prepend # default "append" (after the upstream description)
redact:
rules:
- regex:
<pattern> # scans every text content block. Leading PCRE-style
# inline flags are supported: `(?i)password` →
# case-insensitive, `(?is)` → +dotall, etc.
- field:
<path> # dotted path with * (one level), ** (any depth), [i] (index).
# Auto-parses JSON strings inside text blocks, and
# also processes balanced JSON and Python row spans
# embedded in a narrative envelope (e.g. MongoDB's
# <untrusted-user-data-…> wrapper).
replacement: <string> # default "[REDACTED]"
applyTo: text|all|fields
sqlGuard:
tools: [<tool-name>] # which tool(s) carry a SQL argument
sqlArg: <name> # default "sql"
readOnly: <bool> # default true; rejects mutations in the SQL argument
mode:
allowlist|denylist # default "allowlist" (recommended)
# allowlist: accept only leading read verbs
# (SELECT / WITH / SHOW / EXPLAIN / DESCRIBE /
# VALUES / PRAGMA / TABLE) AND reject any
# embedded DML / DDL keyword or SELECT INTO
# hiding inside a WITH CTE / EXPLAIN ANALYZE.
# denylist: legacy keyword-blacklist; preserved
# for compatibility.
extraReadVerbs:
[<word>, ...] # allowlist mode: dialect-specific
# read verbs your MCP needs on top of the defaults.
extraWriteKeywords: [<word>, ...] # denylist mode only (ignored otherwise).
rateLimit: # array of rules; first matching rule wins
- tool: <name or glob> # exact tool name or "*"-glob (same as `block`)
perMinute: <number> # steady-state rate; bucket starts full at <perMinute>.
# Each matched *tool* gets its own bucket, so one hot
# tool can't starve others that share the same rule.
# `per_minute` (snake_case) is also accepted.
telemetry: # optional; omit entirely for zero-overhead (no-op tracer)
otel:
endpoint: <url> # OTLP HTTP endpoint
# e.g. "http://otel-collector:4318/v1/traces"
serviceName: <string> # optional; default "januscope"
headers: # optional auth / routing headers
Authorization: "Bearer ${OTEL_TOKEN}"
# Install the peer deps only when you enable this:
# npm install @opentelemetry/api @opentelemetry/sdk-trace-base \
# @opentelemetry/exporter-trace-otlp-http \
# @opentelemetry/resources
audit:
sink:
<path|stderr|stdout> # "~" is expanded. Parent directories
# are auto-created. File opens with 0o600
# perms (user-only; matters for logRawArgs).
logRawArgs: <bool> # default false; when false, args are SHA-256 hashed
Environment variables in string values are expanded with ${VAR} or $VAR. Missing variables become empty strings and emit a one-line [januscope] warn: env var 'FOO' is unset, substituted empty string on stderr (once per name). We don't refuse to start, the user may be intentionally testing with undefined vars, but the warning is loud enough that a forgotten $DATABASE_URL at 2am shows up in the logs instead of silently breaking the lens.
Alongside plain ${VAR} substitution, three URI-shaped references are resolved at startup from external secret stores:
#<field> selects one field out of the stored object; if the secret has exactly one field you can omit it. The sync loadConfig() refuses a config with vault references and directs you at the async loadConfigAsync(), the CLI always uses the async path, so januscope --config … handles both cases transparently.
Design note: resolvers fetch at startup only. JanuScope reads the value once, hands the substituted config to the pipeline, and never calls the secret store again for the life of the process. Rotate your secrets; restart JanuScope.
An MCP server is a program that speaks JSON-RPC 2.0 over stdio. JanuScope is also a program that speaks JSON-RPC 2.0 over stdio, it just happens to spawn the real MCP server as a child process and forward messages through a pipeline of overlays.
[AI client] ──stdin/stdout──> [januscope] ──stdin/stdout──> [real MCP server]
▲
rewrites tools/list
short-circuits blocked tools/call
injects schema, scrubs output, logs
No daemon or listening port. Persistent state is limited to configured audit logs and optional first-use approvals; an upstream bridge can maintain its own authentication state. JanuScope lives as long as the client's connection and dies with it. See ARCHITECTURE.md for the full picture.
dbSchema and contextInjection actually workThese are the two overlays that add context to the tool descriptions the LLM sees. The mechanism is the same; what differs is where the text comes from.
Walk-through, dbSchema against Postgres
config.yaml. The dbSchema: block is a recipe (driver, connection string, which schemas to look at, which tools to inject into). The schema text itself is not in the YAML.connectionString and runs a few information_schema queries. It pulls table names, columns, types, foreign keys, and (if includeComments: true) any SQL comments. This takes 50-300 ms and runs once per launch.tools/list to JanuScope. JanuScope forwards it to the real Postgres MCP. The MCP returns its tool list with the standard descriptions.dbSchema overlay rewrites it. For each tool name in injectInto: (typically execute_sql), the schema blob is appended to that tool's description field.list_tables round-trip needed.Schema context explains much of the measured discovery reduction. The original 84% result reset conversation history; see the newer retained-conversation comparisons above.
Walk-through, contextInjection against Linear
contextInjection is the same idea but the operator is the introspector. JanuScope doesn't know how to "introspect" Linear (or Notion, or a filesystem), so the operator supplies the text. Two storage choices:
Inline (text: |). Best for short, hand-curated context that doesn't change often. The text lives directly in the lens's config.yaml:
contextInjection:
injectInto: [list_issues, search_issues, get_issue]
text: |
Active projects: PROJ-A (engineering), PROJ-B (data), PROJ-C (growth).
Issue states: backlog, todo, in_progress, in_review, done, cancelled.
External file (textFile: ./context.md). Best when the context is long, or when an external job (cron, CI, a homegrown script) keeps it fresh. The lens's config.yaml references the file by path:
contextInjection:
injectInto: [list_issues, search_issues, get_issue]
textFile: ./linear-context.md # next to config.yaml
Relative paths resolve against the lens's directory, so a lens that ships with context.md next to its config.yaml works no matter where the operator launches JanuScope from. Absolute paths and ~/... paths also work.
The operator can run a separate cron job that regenerates linear-context.md every hour (or on commits, or whenever they want); JanuScope picks up the new content the next time the proxy starts. JanuScope itself doesn't fetch from Linear; that decoupling is deliberate so the same overlay works for any lens without baking API integrations into the core.
The runtime path is identical to step 4-6 above: on the next tools/list response, JanuScope appends (or prepends, with position: prepend) the text to the description of every tool listed in injectInto.
When to use which
dbSchema for any lens whose target is Postgres / MySQL / SQLite. JanuScope handles the introspection.contextInjection for non-DB lenses where you have a small, stable surface the LLM otherwise has to discover. Good fits: Linear (projects / teams / status enums), Atlassian (projects / spaces), filesystem (directory skeleton for a tight allowed root), Notion (workspace navigation skeleton). Marginal fits: Stripe (mostly direct-by-id usage), GitHub (mostly direct-by-id usage). Bad fits: anything large (full file content, full activity feeds) or volatile (real-time issue counts, recent messages).dbSchema for the SQL tool and contextInjection to add a glossary of enum values to a separate non-SQL tool.Because JanuScope is a stdio proxy spawned as a child of your MCP client, its lifecycle is the client's lifecycle, there is no daemon and no /health endpoint. That is deliberate: no listening port, with audit files and optional approval state kept locally. It also means the usual process-supervisor patterns don't apply directly. Here's the shape of the failure modes and what to do:
Uncaught exception or fatal unhandled rejection. JanuScope observes Node's uncaughtExceptionMonitor event and writes a runtime diagnostic. It does not install handlers that suppress Node's normal fatal behavior. With Node's default error policy, the process exits nonzero and the MCP connection closes. The default diagnostic writes synchronously before exit:
[januscope:runtime] 2026-01-01T… error: unhandledRejection {"message":"…","stack":"…"}
An embedding application's handlers, capture callback and Node rejection flags retain their own behavior. A handled error therefore does not automatically disconnect that application's MCP sessions. JanuScope removes its monitor when a session finishes. When Node emits its exit event, each active bridge sends SIGKILL to its owned direct target. Forced OS termination, abort modes such as --abort-on-uncaught-exception, and failures thrown inside host fatal-error handlers can skip these cleanup hooks. This does not promise cleanup of independently daemonized descendants or a final asynchronous audit/telemetry flush. Runtime diagnostics can contain the original error message and stack and should be treated as sensitive.
The wrapped MCP crashes or a connection is stopped. Graceful shutdown first awaits overlay stop hooks. After those hooks settle, the transport closes target stdin and schedules SIGTERM after two seconds, SIGKILL after five seconds, and a ten-second wait deadline. A hanging stop hook can delay this sequence; these are not per-request timeouts or a general hang detector. Reconnection depends on the MCP client. Reconnect or restart the server from the client to launch a fresh process pair.
Stream errors. Errors from client stdin or target stdout are logged as warn; those handlers do not themselves stop the bridge. EOF, stream end and target exit initiate the shutdown described above. Writes to closed client output or target input can emit asynchronous EPIPE errors. Under Node's default error policy, these unhandled errors follow the fatal diagnostic and nonzero-exit path above. A stream error alone does not guarantee a graceful exit or resolution of done.
Running JanuScope under a supervisor. If you embed JanuScope in a long-running sidecar (e.g. a custom gateway that keeps one stdio bridge per LLM session), supervise it the same way you would any Node process: systemd with Restart=on-failure, or Docker's restart: unless-stopped. The stderr lines above give the supervisor enough to distinguish crashes from clean exits.
No /health, on purpose. A stdio-only proxy that added an HTTP health endpoint would defeat the "no open listeners" posture. For liveness, use tools/list on an initialized MCP connection, or rely on the client's connection state. The optional npx -y januscope check --config <name-or-path> diagnostic performs initialization and discovery for a separate test connection.
Remote presets currently use a separate stdio-to-HTTP bridge. Native HTTP transport is not implemented in the proxy.
When a lens enables audit:, JanuScope writes a structured JSONL log of tracked call outcomes plus startup and shutdown events, one JSON object per line. Write errors are logged and do not stop forwarding, so monitor the sink and runtime diagnostics when log completeness matters.
audit:
sink: ~/mcp-audit-postgres.jsonl # file path, or "stdout" / "stderr"
logRawArgs: false # default; true = include un-hashed arguments
~ expands to your home directory.0o600 (user-read/write only). Existing files keep their current mode, if you later flip logRawArgs: true, rotate to a fresh sink so the tighter perms apply.stderr sends audit lines to the diagnostic stream. stdout is supported by the overlay but shares the MCP protocol pipe in the CLI; use a file or stderr for an MCP-client connection.Every record has these base fields:
Plus event-specific fields:
startup, written once when the pipeline initialises:
{ "ts": "2026-04-18T12:00:00.000Z", "event": "startup", "sink": "~/mcp-audit.jsonl" }
tools/call, written once when a tracked response arrives, with request metadata and outcome in the same record. A missing response may instead be recorded as timeout or orphaned:
{
"ts": "2026-04-18T12:00:01.456Z",
"event": "tools/call",
"tool": "query",
"id": 42,
"args_hash": "sha256:0123456789abcdef",
"status": "ok",
"duration_ms": 333,
"result_bytes": 2048
}
This record is illustrative. IDs, hash, timing, and byte count depend on the actual call.
Fields:
Local MCP proxy for tool restrictions, response redaction, audit logs, and database schema context.
{ "mcpServers": { "io.github.giancarloerra-januscope": { "command": "npx", "args": [ "-y", "[email protected]" ], "env": { "JANUSCOPE_USER": "<value>", "JANUSCOPE_TEAM": "<value>", "JANUSCOPE_SESSION": "<value>" } } }}Herramientas
Aún no hay lista de herramientas. Los servidores remotos la reciben tras la primera comprobación de estado.
Preguntas frecuentes
Copia en esta página la configuración stdio para Cursor, Claude Desktop, Claude Code o VS Code. Usa la instalación del publicador, no un proxy de PlugMCP.
El publicador indica estas variables de entorno: JANUSCOPE_USER, JANUSCOPE_TEAM, JANUSCOPE_SESSION.
Januscope se ejecuta localmente por stdio.
Última comprobación: 2026-11-14. Local MCP proxy for tool restrictions, response redaction, audit logs, and database schema context.