cc-code-review
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| CLAUDE_CODE_PATH | No | Path to the claude executable or cli.js to use (auto-detected by default; shell shims such as .cmd/.bat/.ps1 are not accepted). | |
| CLAUDE_REVIEW_MODEL | No | Default model to use for reviews. | |
| CLAUDE_REVIEW_LOG_DIR | No | Directory for run logs (logs contain excerpts of reviewed code; on Windows there are no POSIX permission bits, so point this at a directory with your own ACLs for stricter isolation). | %TEMP%/cc-code-review |
| CLAUDE_REVIEW_SHELL_GUARD | No | Set to 0 to disable the PreToolUse shell chained-command interception (only for debugging false positives; not recommended, the server will still fail closed). | 1 |
| CLAUDE_REVIEW_TIMEOUT_MIN | No | Default review timeout in minutes. | 20 |
| CLAUDE_REVIEW_MAX_BUDGET_USD | No | Default API spend cap in USD (0 disables the cap). | 5 |
| CLAUDE_REVIEW_ALLOW_UNGUARDED | No | Set to 1 to allow reviews to run when the PreToolUse hook is not active or its self-test fails. By default the server refuses to run in that case. | |
| CLAUDE_REVIEW_AUTO_INSTALL_SKILL | No | Whether to automatically install the bundled code-review skill when it is missing. | 1 |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Run Claude Code's code-review skill on a repository and get structured findings back. Typical flow: claude_code_review(repoPath, scope) returns a jobId after a fast preflight -> poll claude_review_status(jobId) until it returns the report -> fix the code -> run claude_code_review again to verify, or claude_review_followup(sessionId, message) to ask the reviewer something. Pass waitForResult=true only when the client can hold the call for the whole review.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| claude_review_envA | Check the Claude Code CLI installation, the code-review skill, the auth/gateway setup, and helper tools (git/gh). Set installSkill=true to (re)install the bundled code-review skill into ~/.claude/skills/code-review (never overwrites a user-managed skill there). Diagnostic contract: the verdict (OK / NOT_FOUND / MISCONFIGURED) is reported in the text and in structuredContent; isError is reserved for a failure of this diagnostic itself. |
| claude_code_reviewA | Run Claude Code's code-review skill (a diff-sized fleet of parallel reviewer subagents, 0-100 confidence scoring, keeps findings >=75). Every finished reviewer subagent's report is captured the moment it lands — appended to a durable JSONL file and visible via claude_review_status while the run continues — and an agent killed mid-flight (e.g. a 429) is recovered from its output_file transcript, so a rate limit never wastes the work already paid for. on a repository and return the markdown report. Costs Claude API credits; a run typically takes 2-15 minutes. Afterwards, fix the code and re-run to verify, or use claude_review_followup with the returned session_id. |
| claude_review_statusA | Poll a background review started by claude_code_review (real reviews take 2-15 minutes, so the async call returns a jobId). With jobId: returns elapsed time while running, or the full report/result when finished. With cancel=true: aborts the run and stops billing (requires a jobId). Without jobId: lists this server process's jobs. Retention: the 10 most recent finished jobs are kept in this server process; a restart clears them all. |
| claude_review_followupA | Continue the conversation of a previous claude_code_review run (same reviewer context, e.g. '为什么第 2 条是问题？' or 're-check after my fix'). Pass the session_id returned by claude_code_review. Async by default (returns a jobId after a fast preflight; poll claude_review_status) — waitForResult=true blocks synchronously. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 4 tools
Each tool covers a distinct phase of a single workflow: environment/diagnostic setup, launching a review, polling/cancelling a job, and continuing a prior session. The only potential overlap (claude_code_review vs claude_review_followup) is clearly separated by 'start new run' vs 'pass session_id to continue'. Descriptions explicitly disambiguate async vs sync behavior.
All names share a claude_ prefix and a noun-action feel (review_env, code_review, review_status, review_followup), making the family obvious. Minor deviation: claude_code_review inserts 'code' where others use claude_review_*, so the prefix isn't perfectly uniform.
Four tools is a tight, well-scoped surface for a code-review runner: setup check, run, monitor, follow-up. Each tool maps to a necessary step and none feels redundant or missing for the stated purpose.
The lifecycle is essentially covered: diagnose/install skill, run review, poll/cancel, and follow up with the same reviewer context. Minor gaps: no dedicated tool to retrieve or list past reports/JSONL transcripts, and no explicit way to configure review parameters (scope, thresholds) beyond defaults.