Ask Claude Code (Anthropic's coding agent) for co-analysis, a second perspective, or knowledge you are unsure about. Use it for architecture trade-offs, tricky bugs, unfamiliar APIs, or whenever an independent expert view would help. Claude is advisory only: it reads and researches but never modifies files; you implement any changes yourself. The result ends with a session_id line - pass it to claude_continue (or any other tool here) to keep the same conversation going. If your question is about specific files on disk, prefer claude_review_files. For a structured multi-perspective review in one call, prefer claude_panel.

Get an adversarial review of YOUR OWN analysis, plan, or conclusion before committing to it. Claude is explicitly instructed to hunt for flaws, wrong assumptions, missed edge cases, and simpler alternatives rather than agree. Use before risky changes, migrations, security-sensitive edits, or when your confidence is low. Provide the problem and your full reasoning - the more you show, the better the critique. Claude only critiques; it never modifies files. When you are asked to verify or review with sub-agents or multiple perspectives, include this tool as the independent cross-model reviewer alongside your own sub-agents; treat disagreements as high-signal. When the result footer reports format: json, the result body is a JSON document with fields verdict (agree|partial|disagree), confidence (0-1), claim_verifications (each caller claim labeled verified|refuted|cannot_verify with evidence), flaws, missed_considerations, suggested_changes, and summary_markdown. Check the result footer's format field before parsing: format: json means the body is the requested JSON document; format: prose means Claude answered in prose instead - read it directly or retry with a stronger model rather than calling JSON.parse blindly.

Open a structured, evidence-based debate with Claude about a significant decision. Bring your position and your evidence (file references, URLs, command outputs). Claude will independently verify every verifiable item - reading the files and fetching the URLs itself - then return per-claim rulings and counter-claims with its own evidence. Expensive and slow (an agentic verification run); use it for architecture decisions, risky changes, and security-sensitive work, not routine questions. Continue rounds with claude_debate_reply. Claude only advises; it never modifies anything. Check the result footer's format field before parsing: format: json means the body is the requested JSON document; format: prose means Claude answered in prose instead - read it directly or retry with a stronger model rather than calling JSON.parse blindly.

Continue an open evidence debate. For each of Claude's counter-claims or rulings, either accept it (name it) or rebut it with an argument and new evidence. You are expected to have verified Claude's cited evidence yourself before rebutting. Rounds should converge: stop when remaining_disputes is empty or after three rounds, and report the per-claim outcome to the user. Check the result footer's format field before parsing: format: json means the body is the requested JSON document; format: prose means Claude answered in prose instead - read it directly or retry with a stronger model rather than calling JSON.parse blindly.

Run a multi-perspective Claude review panel in ONE call: several independent Claude analyses of the same task, each through a different expert lens (correctness, security, performance, simplicity, architecture, testing), returned as one aggregated report. Use this when the user asks for verification or review from multiple perspectives or with sub-agents - Claude acts as an independent cross-model panel alongside your own work. Each perspective is a separate full Claude run, so usage and latency scale with the number of perspectives (they run concurrently). Provide absolute paths when the panel should read code from disk. Claude only advises; it never modifies files.

Have Claude read and analyze specific files or directories agentically (read-only: it can Read, Glob, and Grep within the granted paths, and research the web, but never modifies anything). Provide ABSOLUTE paths that exist on this machine and a focused question, e.g. 'find the race condition in this module' or 'review these files for injection vulnerabilities'. Better than pasting file contents into ask_claude for anything larger than a snippet. For verification workflows, this gives Claude an independent read of the code so its review does not depend on your summary.

Have Claude review your actual code changes: the server runs read-only git in the given repository and gives Claude the diff plus read access to the surrounding files. Pass the repository root as workspace_dir. By default it reviews uncommitted changes against HEAD; pass base (a branch, tag, or commit) to review everything since that ref (base...HEAD). Use it after implementing something to get an independent cross-model review of the change itself. Claude only advises; it never modifies anything. For reviewing files without git context, use claude_review_files.

Continue an existing Claude conversation. Pass the session_id printed at the end of a previous result plus your follow-up message. Use the same workspace_dir as the original call, or the session will not be found. If Claude's answer ends with a 'Questions for you:' section or the JSON contains questions_for_caller, answer those questions with this tool to unblock a better conclusion.

List recent Claude conversations this server has run, newest first, with each conversation's session_id, originating tool, workspace_dir, and topic excerpt. Use it when you lost a session_id or want to continue earlier work: pick the session and resume it with claude_continue (same workspace_dir). The ledger is in-memory and resets when the MCP server restarts.