Loading ToolBench
Local pre-load inspection for GGUF artifacts and llama.cpp runtimes. Exposes security inspection capabilities as MCP tools for agentic use.
Cancerbero defines 5 tools with complete JSON schemas and substantive descriptions (avg 180 chars). All tools are read-only, reducing risk. Naming is clear and verb-driven (inspect, check, scan, list). Schemas include proper type constraints, enums, patterns, and parameter descriptions. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance or actionable error messages visible. Tool composition is sound (each does one thing), but descriptions lack dependency hints and prerequisites.
Read the GGUF metadata and tensor descriptors of a single artifact WITHOUT running template or companion analysis. Useful as a fast first step before deciding whether the full check is warranted. Returns the same `ArtifactFacts` payload as `cancerbero_inspect` but with no findings.
Run the chat-template analysis on a single template string without needing a GGUF file. Reports AST findings, SSTI gadgets, exfiltration URLs, and prompt-injection patterns. Use this when the agent already has the template text in context and wants to know whether it is safe to render.
Audit a model repository directory for companion files (config.json, tokenizer.model, etc.) that may contain malicious code or unsafe patterns. Returns findings for each companion file discovered.
Run the full Cancerbero check on one or more targets. Returns the verdict, the list of findings, optional SHA-256 digests, and the per-artifact / per-runtime facts. This is the default tool to call when an agent needs to decide whether a GGUF artifact and an optional llama.cpp runtime are safe to load.
List all security advisories in the current knowledge bundle. Returns the advisory rules that Cancerbero uses to evaluate runtime and artifact safety. Useful for understanding what the current installation can detect.
Output schemas not documented. LLMs cannot predict response structure (e.g., what fields are in ArtifactFacts, what does a finding object contain). This forces agents to guess and wastes tokens on exploratory calls.
No error handling guidance. Tools return findings/verdicts but do not document what errors are possible (e.g., file not found, invalid GGUF, permission denied) or how to recover. Agents cannot self-correct.
Missing dependency hints. cancerbero_artifact_facts is described as 'fast first step before deciding whether full check is warranted' but does not explain when to call it vs cancerbero_inspect, or what the agent should do with the facts returned.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Parameter 'runtime_version' lacks format guidance. Description says 'e.g. b8146 or 0.2.72' but does not specify the format constraint (regex, length, allowed characters). LLMs may pass invalid versions.
STDIO transport only. Server is not remotely accessible and cannot be used by hosted MCP clients (Claude.ai, Cursor Cloud, etc.). Limits deployment scenarios.