MCP server for CVE/config remediation orchestration: SSH inventory, OSV/NVD lookups, CIS/DISA STIG-mapped Ansible playbook suggestions.
prescryb demonstrates solid tool design with clear naming (verb_noun pattern), comprehensive descriptions (avg 250+ chars), and well-structured JSON schemas. All 9 tools have explicit input schemas with typed parameters and descriptions. However, output schemas are not formally documented in the code, responses are returned as dicts without schema declarations. Error handling is minimal; tools lack recovery guidance and actionable error messages. Security is strong (no credential parameters, SSH key handling via agent/config). Composition is excellent: tools chain logically (inventory → check_cves → fetch_advisory → map_compliance → lookup_cce → generate_playbook). Tool names are action-oriented and unambiguous. Parameter descriptions are detailed and include constraints (e.g., 'trust_unknown_host' rationale, CCE platform limitations). Average tool score: 72/100.
Match installed package versions against known CVEs via OSV.dev. `system` and `packages` are the objects returned by inventory_host (or a filtered subset of `packages`). Matches are enriched with `epss_score`/`epss_percentile` (see fetch_epss); a FIRST.org outage degrades to unset EPSS fields plus a `warning`, not a failed match.
Fetch the current, authoritative NVD record for a specific CVE ID. Use this to get an up-to-date description, CVSS score/severity, CWE weakness classification, and reference links for a CVE surfaced by check_cves (or one you already know about), rather than relying on potentially stale training data.
Fetch EPSS scores for CVE IDs not already scored by check_cves. Unscored CVE IDs (too new, reserved, rejected) are absent from `scores`, not an error.
Generate a suggest-only Ansible playbook for human review. Produces an Ansible playbook based on CVE matches and compliance findings. This tool never applies changes to the target host itself.
SSH into `host` and inventory installed packages. Auth uses ~/.ssh/config, an SSH agent, and default identity files - the same as running `ssh host` yourself. No password argument is accepted: credentials must never flow through MCP tool-call arguments. Unknown host keys are rejected unless trust_unknown_host=True; prefer connecting with `ssh host` once yourself to pin the key instead. `hostname`/`identity_file` override the resolved address and key path without editing ~/.ssh/config - handy for e.g. a local molecule/vagrant instance. Only a path is passed, never key contents.
Output schemas not formally documented. Tools return dicts (e.g., inventory_host returns {system, packages, package_count}) but no schema declarations guide LLM on expected fields, types, or downstream chaining. LLMs must infer structure from examples.
Error handling lacks recovery guidance. Tools do not return actionable error messages (e.g., 'CVE not found in NVD. Try fetch_epss() for EPSS-only scores' or 'SSH connection failed: unknown host key. Set trust_unknown_host=true or run ssh host manually first'). Raw exceptions provide no LLM guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
List platform names lookup_cce can query (e.g. 'rhel8', 'firefox'). Sourced live from https://github.com/konstruktoid/cce-web, the community JSON conversion of NIST's CCE spreadsheets. Coverage of this project's target distros is thin: only RHEL-family ('rhel6'/'rhel7'/'rhel8') and SUSE ('SLES12-DISA-STIG'/'SLES15-DISA-STIG'/'SLES15-PCI-DSS') are usable - Debian, Ubuntu, Alpine, and Arch have no CCE data upstream.
Look up NIST Common Configuration Enumeration entries for a platform. `target`: free text naming a platform, resolved against the export names published by https://github.com/konstruktoid/cce-web (NIST itself only publishes CCE as spreadsheets). For a host from inventory_host, use 'rhel<major version>' for any RHEL-family distro_id (rhel, almalinux, rocky - CCE only tracks the upstream RHEL number) or 'SLES<major version>-DISA-STIG' for suse; Debian, Ubuntu, Alpine, and Arch have no CCE coverage upstream at all. Call list_cce_targets to see every published platform (also covers e.g. 'firefox', 'apache-httpd2.2', 'win2k8r2'). Pass `keyword` (matched against title/description/rationale/config group) or `cce_id` (exact ID, e.g. 'CCE-80876-6') to filter results. Without either, only the platform's config-group categories and entry count are returned - dumping an entire platform (hundreds of entries) isn't useful; narrow with a keyword first.
Map a free-text topic to CIS Benchmark / DISA STIG topic areas and ATT&CK. `area` examples: 'ssh', 'sudo', 'kernel modules', 'password policy'. If a matching role is found in the konstruktoid.hardening GitHub repo, it is returned too. Also returns the MITRE ATT&CK techniques (and, where defined, the ATT&CK mitigation) that hardening this area addresses. Only topic-area mapping is returned for CIS/DISA STIG, never fabricated specific rule IDs (e.g. "CIS 5.2.1") - those require the licensed benchmark text. Consult the referenced role's own docs/tags for exactly what it covers. ATT&CK technique/mitigation IDs, by contrast, are MITRE's own public catalog (attack.mitre.org), so they're cited directly.
Semantic (RAG-style) search over local reference documents. Searches Markdown, text, and PDF files from a local directory (LOCAL_DOCS_DIR, default local_docs/) holding material the network-backed tools can't see: runbooks, policy, prior remediation notes. Embeddings are computed locally with a sentence-transformers model.
generate_playbook description is vague ('suggest-only Ansible playbook for human review'). Does not explain what 'findings' structure is expected, how it relates to check_cves/map_compliance output, or what the playbook output format is (YAML string? dict? file path?).
lookup_cce 'keyword' and 'cce_id' parameters are mutually exclusive but not explicitly documented as such. Description says 'Pass keyword OR cce_id' but does not state what happens if both are provided or neither is provided.
search_local_docs 'limit' parameter lacks bounds. No min/max specified; LLM could pass limit=999999, causing memory/performance issues. Should constrain to e.g. 1 - 100.