A side-by-side look at the Graft and PatchWarden MCP servers: transport, auth, tools, quality signals, and the config to install either one in Claude, Cursor, Windsurf or VS Code.
Verdict: choose Graft for coding agents work as a local process; choose PatchWarden for coding agents as a local process. You can also run both side by side in one mcpServers block.
GRAFT_ROOT_PATH.safe_read, graft_edit, file_outline.CONTROL_PLANE_API_KEY, PATCHWARDEN_OWNER_TOKEN.save_plan, get_plan, health_check.Rows where the two servers differ are marked.
| Spec | ||
|---|---|---|
| Category | Coding Agents | Coding Agents |
| Quality signal | 65/100 · Great | 64/100 · Good |
| Install command | Confirmed | Confirmed |
| Transport | Local subprocess (stdio) | Local subprocess (stdio) |
| Auth | No auth required | Not declared |
| Pricing | Free | Not declared |
| Required env vars | GRAFT_ROOT_PATH | CONTROL_PLANE_API_KEYPATCHWARDEN_OWNER_TOKEN |
| Verified owner | Community listing | Community listing |
| GitHub stars | 17 | 6 |
| Directory activity | 1 views 0 installs 0 upvotes | 1 views 0 installs 0 upvotes |
| Tools | 34 | 30 |
| Full listing | View Graft | View PatchWarden |
safe_readPolicy-enforced file read. Returns full content for small files, structural outline with jump table for large files, or refusal with reason code for banned files. Detects re-reads and returns cached outlines or structural diffs.graft_editPerform one governed exact string replacement in a repo file. Requires path, old_string, and new_string. Refuses missing or ambiguous old_string matches and policy-denied paths.file_outlineStructural skeleton of a file — function signatures, class shapes, exports. Includes a jump table mapping each symbol to its line range for targeted read_range follow-ups.read_rangeRead a bounded range of lines from a file. Maximum 250 lines. Use jump table entries from file_outline or safe_read to target specific symbols.changed_sinceCheck if a file changed since it was last read. Returns structural diff (added/removed/changed symbols) or 'unchanged'. Peek mode by default; pass consume: true to update the observation cache.graft_diffStructural diff between two git refs. Shows added, removed, and changed symbols per file — not line hunks. Defaults to working tree vs HEAD.run_captureExecute a shell command and return the last N lines of output (default 60). Responses include an explicit policy boundary. Persisted logs can be disabled and obvious secrets are redacted before writing to .graft/logs/capture.log.state_saveSave session working state (max 8 KB). Use for session bookmarks: current task, files modified, next planned actions.state_loadLoad previously saved session state. Returns null if no state has been saved.doctorRuntime health check. Shows project root, parser status, active thresholds, session depth, message count, and burden summary.activity_viewInspect recent bounded local artifact history for the active workspace, anchored to the current commit when possible.causal_statusInspect the active causal workspace and persisted local-history posture.save_planSave an execution plan — ChatGPT writes the plan, PatchWarden stores it for local agent execution. Supports plan_ref to load a plan file already placed inside .patchwarden/plans.get_planRead a saved plan by its plan_id.health_checkCheck MCP catalog consistency, watcher freshness/supervisor state, workspace readiness, and configured agents. Use detail=self_diagnostic for expanded read-only evidence.list_agentsList configured local agents and check whether each executable currently exists. This does not start an agent or contact its model provider.create_taskCreate a repo-scoped task from exactly one source. For ChatGPT, prefer the guarded inspect_only, feature_small, or fix_tests template when it fits. A stale watcher preserves the task but returns execution_blocked and directs the client to health_check; otherwise call wait_for_task until terminal. Use execution_mode=assess_only to pre-assess risk and get an assessment_id without creating a task; then invoke the returned next_tool_call using only execution_mode=execute and the full assessment_id.get_task_statusCheck task status, execution phase, watcher health, pending reason, current command, timeout, and change evidence.get_resultRead result.md, or return structured availability and watcher evidence while the task is not terminal.get_result_jsonRead the structured result.json for deterministic task acceptance.get_diffRead task diff evidence, or return structured availability and watcher evidence while it is not ready.get_test_logRead test.log, or return structured availability and watcher evidence while it is not ready.list_workspaceList files and directories within the workspace (sensitive files excluded).read_workspace_fileRead a file within the workspace. Sensitive files (secrets, keys, tokens) are blocked. In Direct mode (with session_id), reads are scoped to the session's repo_path and return sha256.Paste either block (or both) into your client config, e.g. claude_desktop_config.json or ~/.cursor/mcp.json. We only show a config when the install command is confirmed.
{
"mcpServers": {
"flyingrobots-graft": {
"command": "npx",
"args": [
"-y",
"@flyingrobots/graft"
],
"env": {
"GRAFT_ROOT_PATH": "YOUR_GRAFT_ROOT_PATH_HERE"
}
}
}
}
{
"mcpServers": {
"jiezeng2004-design-patchwarden": {
"command": "npx",
"args": [
"-y",
"patchwarden@0.6.1"
],
"env": {
"CONTROL_PLANE_API_KEY": "YOUR_CONTROL_PLANE_API_KEY_HERE",
"PATCHWARDEN_OWNER_TOKEN": "YOUR_PATCHWARDEN_OWNER_TOKEN_HERE"
}
}
}
}