MCP Server for Sliver C2 adversary emulation framework. Provides tools for managing implants, sessions, beacons, listeners, and more.
This server presents a critical architectural problem: tool definitions are embedded in a Python subprocess script as embedded string literals (lines in index.ts showing `const pythonScript = ...`), not declared as explicit MCP Tool objects. This violates the principle that tools must be registered with the SDK. While the TypeScript code shows parameter schemas in the tool list (e.g., session_id, host, port), I cannot verify these are actually wired into the MCP SDK's CallToolRequestSchema registration. The descriptions are present but generic (10 - 50 chars for many tools, well below the 50 - 200 char baseline for LLM-optimized descriptions). Parameter descriptions are minimal ('The ID of the session' is vague; no guidance on format or constraints). No output schemas are documented anywhere in the code. Error handling is absent, the Python bridge returns JSON with 'error' fields, but there's no structured error recovery guidance for LLMs. Schemas exist in the tool list but appear to be documentation artifacts rather than enforced constraints. Security is concerning: the Python script is injected via string interpolation, creating a supply-chain attack vector if config paths are manipulated. Overall, this reads like a proof-of-concept, not production-grade tooling.
List all beacons
Generate an implant with specified options. Note: Direct implant generation requires more complex setup - use Sliver console for full functionality.
List all implant builds
List all implant profiles
Interact with a specific beacon by ID
Interact with a specific session by ID
List all jobs (listeners and other background tasks)
Kill/stop a job (listener or background task)
Tool definitions are embedded in Python subprocess script strings, not registered explicitly with MCP SDK. Cannot verify actual CallToolRequestSchema binding. This makes tools functionally invisible to the protocol.
No output schemas documented. Python code constructs JSON responses, but there is no schema documentation for LLMs to understand the structure of results (e.g., what fields does 'operators' return? what are their types?). This forces LLMs to infer structure from examples.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Terminate a session
List all operators connected to the Sliver server
Change directory on a session
List directory contents on a session
Create a directory on a session
Get current working directory on a session
List all active sessions
Start a DNS listener
Start an HTTP listener
Start an HTTPS listener
Start an mTLS listener
Start a WireGuard listener
Get Sliver server version information
Tool descriptions are generic and under 50 characters for most tools (e.g., 'List all operators connected to the Sliver server' is 52 chars but lacks context on WHEN to use it or what to do with results). Baseline for LLM-optimized descriptions is 50 - 200 chars with explicit intent. Current descriptions fall short of LLM guidance standards.
Parameter descriptions are minimal. 'session_id: The ID of the session to interact with' provides no guidance on format, constraints, or how to obtain a session_id if the LLM does not have one. Should include: 'Call sessions() first to get a valid session_id. Format: alphanumeric string.'
No error recovery guidance. If a tool call fails (e.g., 'operator config file not found'), the Python script returns a JSON error object, but there is no structured error classification or next-step guidance for the LLM. Should return: 'Resource not found: operator config file. Action: Check --operator-config flag or SLIVER_OPERATOR_CONFIG env var.'
No pagination support documented or implemented. Tools like 'sessions', 'operators', 'beacons', 'jobs', and 'implant_builds' could return large result sets with no limit parameter or pagination tokens, risking context window exhaustion.
generate() tool description notes 'Direct implant generation requires more complex setup', signaling incompleteness. The tool should either fully work or be removed. Incomplete tools confuse LLMs.
Python script injection via string interpolation creates a supply-chain attack vector. Config paths and parameters are embedded directly into a Python script string. If an attacker controls operatorConfigPath or any parameter, they can inject Python code. Mitigate by using subprocess argument arrays and JSON serialization.
No tool annotations. Tools that modify state (kill_session, kill_job, start_*_listener, session_mkdir, session_cd) are not marked as destructiveHint or non-idempotent. Without annotations, LLMs cannot reason about retry safety.