Anticells Red MCP
Saved research reports
Retrieve, export, back up, and manage saved research evidence.
Research tools automatically save evidence in a separate SQLite database. Each MCP process creates a session; each assessment, mutation call, or background loop creates a run. Replies include run_id, session_id, and persistence. No database server, browser service, or PDF dependency is needed.
execute_attack_template creates one run for one selected template. Its run includes the stable template ID, namespace, record revision, and catalog revision. Execution or a compatibility skip produces one evidence record; a configuration error produces a failed run with no target attempt. Its results also enter current-session history, with local mutation and skipped scenarios excluded from target-attempt totals. Export the returned run_id using the same workflow below.
Storage and setup
get_fingerprint_profile saves each probe once and stores the completed profile under the run's fingerprint_profile. A saved summary view includes the profile; detailed views and exports additionally include probe evidence. This retrieval needs neither the target nor the catalog. See profile generation and retrieval.
Startup creates reports.sqlite3, applies versioned report migrations, and prepares an exports directory. The default directory is <state directory>/reports, where the state directory is the same directory used for the default catalog. On Windows it defaults to %USERPROFILE%/AppData/Local/anticells_red_mcp/catalog/reports; on Linux, ~/.local/share/anticells_red_mcp/catalog/reports; on macOS, ~/Library/Application Support/anticells_red_mcp/catalog/reports. Use ANTICELLS_RED_MCP_STATE_DIR for an explicit shared state location on any platform.
Set ANTICELLS_RED_MCP_REPORTS_DIR to choose a different directory. ANTICELLS_RED_MCP_STATE_DIR changes the shared default state directory. Setting ANTICELLS_RED_MCP_DATABASE_PATH alone changes only the catalog file. Use a local disk outside a synchronized source checkout. Environment files are loaded only when explicitly selected with ANTICELLS_RED_MCP_ENV_FILE or the target profile configuration.
The report database stores complete prompts, generated conversations and tool schemas, normalized reply text, proposed tool calls, outcomes, original result timestamps, execution status, catalog revision, and safe operation settings. Authentication headers, environment snapshots, rendered custom request bodies, and configured credential values are excluded or redacted. Endpoint labels omit query strings. Full target evidence can itself contain sensitive information: files stay local, with no automatic sharing or expiration. Unix directories/files use 0700/0600; Windows inherits permissions from the selected user directory.
Use through a coding agent
These examples are arguments to get_attack_report, not shell commands. The server exposes twelve tools, including read-only template discovery.
List saved runs in the current MCP session:
{"action": "list", "limit": 20, "offset": 0}
The result includes total, next_offset, and scope. To list a previous session, supply its explicit ID:
{"action": "list", "session_id": "<session UUID>", "limit": 20}
Use the reports CLI's list command to discover saved runs across all sessions. An explicit run_id can retrieve an older run directly. View a run's statistics and up to 100 evidence records:
{"action": "view", "run_id": "<run UUID>", "format": "detailed"}
Use session_id instead of run_id to combine all runs in a session. summary returns statistics/provenance; json includes up to 10 evidence records; detailed includes up to 100. evidence_truncated indicates omitted inline records. Export for complete evidence:
{"action": "export", "run_id": "<run UUID>", "export_format": "markdown"}
export_format accepts json, markdown, or html and defaults to json. With neither selector, export chooses the latest completed run in this MCP session. Explicit IDs can select old sessions or active work. An active export is labeled partial; it is an immutable snapshot and does not grow as the run continues. Every export creates a new file with a unique name.
Export replies include export_id, filename, absolute local path, MIME type, byte count, SHA-256 checksum, and an MCP resource link. The agent can attach/copy the local file or retrieve the resource. Whether a download button appears depends on the coding agent; MCP does not prescribe that UI.
Resource listings show exports from the current MCP session. Registered export IDs support whole-file reads up to 1 MiB. For larger files, or agents without filesystem access:
{"action": "read", "export_id": "<export UUID>", "offset": 0, "length": 65536}
Decode data from base64, append the bytes in order, and request next_offset until eof is true. Check the reconstructed byte count and SHA-256. Each chunk is at most 64 KiB. The server accepts registered export IDs only, never filesystem paths.
Existing calls such as {"format":"summary"} keep their original current-session view and accounting. Saved-run/session selections additionally include standalone fingerprint probes and local mutation candidates. Failures and skipped scenarios are archived alongside successes. Local mutation and skipped scenarios do not count as target attempts. Confidence remains a heuristic label and is never converted into a vulnerability severity.
CLI, backup, and cleanup
The CLI examples below assume the project virtual environment is activated. Otherwise use .venv/bin/anticells_red_mcp-reports on macOS/Linux or .venv\Scripts\anticells_red_mcp-reports.exe on Windows, from the repository root.
The separate CLI reads reports without loading a prompt catalog or contacting a target:
anticells_red_mcp-reports list --limit 20
anticells_red_mcp-reports show --run-id <run-uuid>
anticells_red_mcp-reports export --session-id <session-uuid> --format html
anticells_red_mcp-reports export --run-id <run-uuid> --format json --output ./assessment.json
anticells_red_mcp-reports delete --run-id <run-uuid>
anticells_red_mcp-reports delete --session-id <session-uuid>
Use --reports-dir /absolute/private/path before the command to override the directory. For example, anticells_red_mcp-reports --reports-dir /absolute/private/path list inspects that store. CLI storage settings come from ANTICELLS_RED_MCP_ENV_FILE and the process environment; the profiles file's env_files are not applied to its directory selection. Pass the report directory explicitly if the server selects it there. python -m anticells_red_mcp.reports.cli provides the same commands. Explicit output copies refuse to overwrite an existing file. Deletion removes selected evidence and dependent exports, including combined exports containing a deleted run. Active runs and active sessions cannot be deleted. No automatic cleanup occurs.
For a complete portable backup, stop all MCP processes using this directory, then copy the entire report directory, including reports.sqlite3 and exports. Restore it into a private local directory and set ANTICELLS_RED_MCP_REPORTS_DIR accordingly. Registered export paths are reconstructed from their IDs and filenames, so moving the directory is supported. Do not copy a live SQLite file directly; use SQLite's backup API if an online database backup is required. For a consistent database-plus-export backup, stopping writers is recommended. The catalog database is backed up separately.
Recovery and evidence format
Execution statuses are running, completed, failed, stopped, and interrupted. Failed target attempts do not necessarily make the whole run fail. Configuration errors are run failures with an error label and do not become target attempts. Session lock files use OS locks to distinguish a dead process from another live coding-agent session. Abandoned runs become interrupted and are never automatically resumed.
Saving errors do not stop testing. A persistence warning appears in tool replies and background reports; pending records remain in memory, retry every five seconds, and are retried during orderly shutdown. Stable result IDs prevent duplicate inserts after recovery. If the process exits before recovery, unsaved evidence can be lost. A recovered interrupted run reports that persistence completeness is unknown. Exports disclose incomplete persistence; export failures return report_error and do not claim a saved file.
Saved report access remains available when catalog loading fails. Research calls continue to report the catalog error until the catalog is repaired. Report storage errors use static diagnostics that omit database paths, credentials, and evidence.
All formats use the same report document and statistics. HTML escapes target content and contains no scripts or external assets. Markdown displays evidence literally. JSON normally contains a readable report object. For exceptional legacy Unicode that would change during an ordinary JSON round trip, the whole document uses encoded_content with encoding json-utf8-surrogatepass-base64. Base64-decode its data, decode UTF-8 with surrogatepass, then parse JSON to recover the exact original Python values.
Previously lost in-memory reports cannot be recovered. Auxiliary Python-only entry points are not automatically archived. PDF and browser download services are not provided. Native Windows setup and the event-loop fix are included in the current installer; see installation.