GitLab MCP Server vs Superpowers

GitLab MCP Server 30

No x402, MPP or L402 (0). Plan prices are public for Free and Premium, Ultimate is by quote and GitLab Credits are $1 each, but the MCP server has no per-call price (10). GitLab.com Free needs no card, and the MCP server is on every tier (20). A person signs up, and an administrator must enable MCP and a user must authorise in a browser (0). 0 + 10 + 20 + 0 = 30.

Superpowers 60

Self-hosted rule. No payment protocol (0). Free under MIT with nothing to buy (20). Free to use with no card (20). An agent can install it in its own harness with no account or sign-up (20). Model usage is billed by the coding agent's provider, and the enterprise support Prime Radiant mentions has no published price. 60 in all.

GitLab MCP Server 87

GitLab's code is public in gitlab-org/gitlab, with the core under MIT and enterprise functions under GitLab's own licence; we didn't establish which licence covers the MCP server code (20). The privacy statement and a sub-processor page of 63 entries name AI providers and their purposes, but that page shows no update date and links no DPA (20). A dated deprecations and removals page and per-tool version history, with superseded tools kept as callable aliases (20). Sub-processors and their locations are disclosed, and Dedicated customers choose their AWS region (18). 20 + 20 + 20 + 18 = 78.

Superpowers 68

MIT, an OSI licence (30). The README and the source agree on the one outbound request and what it carries, git-ignored scratch files stay in the working tree, and diagnosing-superpowers has a written redaction policy and approval gates. Prime Radiant publishes no privacy policy covering the logo request, which reaches its site with the visitor's address (15 of 30). Breaking changes are dated in the release notes, the legacy slash commands showed deprecation notices in v5.0.0 (9 March 2026) before removal in v5.1.0 (30 April 2026), but no written deprecation policy (12 of 20). Telemetry disclosed in the README with three opt-out variables, read in server.cjs (20). 77 in all.

GitLab MCP Server 85

The tools reference documents every parameter with its type, required flag and enum values, and the server exposes JSON Schema through MCP; we read the reference, not the source (22). docs.gitlab.com/llms.txt is served (10). Each tool says what it does, and superseded tools name their replacement (Use get_merge_request with include: commits), though few say when not to use one (14). Enums for include facets, states, scopes and methods, with documented maximums such as first up to 100 and byte_limit up to 1 MB (13). The docs cover 403 insufficient_scope, 400 for an unknown toolset, 429 headers and troubleshooting for Windows and Claude Code, but give few example responses (11). Per-tool version history and monthly release notes (15). 22 + 10 + 14 + 13 + 11 + 15 = 85.

Superpowers 71

Framework reading, docs a model can follow. Each skill has name and description frontmatter and every harness gets a JSON plugin manifest, but the skills are prose with no typed interface or API reference (12 of 25). All docs are Markdown in the repository, which an agent reads directly; no llms.txt (10). Every description starts with "Use when", using-superpowers sets skill priority and a subagent stop rule, and executing-plans says when to prefer subagent-driven development. When not to use a skill is stated in fewer places (17 of 20). Skills take no typed parameters; the helper scripts have fixed arguments and documented exit codes, such as review-package's exit 3 for an empty range (6 of 15). Worked examples in the debugging and TDD skills, troubleshooting in the OpenCode install guide and a diagnosing skill for failed sessions, with few documented error outputs (11 of 15). Semver releases and a dated RELEASE-NOTES.md (15). 71 in all.

GitLab MCP Server 66

OAuth with dynamic registration and a single mcp scope that carries the user's own GitLab permissions, or fine-grained personal access tokens scoped to the MCP tool resource. The docs say GitLab doesn't verify which client uses a client ID and that REST-created apps don't enforce PKCE (22). Toolset and tool headers narrow the surface, an administrator must enable MCP for the group or instance, 19.4 governance defaults writes to Always Ask in Duo settings and 19.6 adds group rules enforced for every client, in beta (16). The docs leave prompt injection to the user ("use MCP tools only on GitLab objects you trust") (5). GitLab audit events exist on paid tiers, but we didn't read whether MCP calls are recorded (8). security.txt is valid until 31 January 2027 and points at a public HackerOne programme; certifications weren't read this run (15). 22 + 16 + 5 + 8 + 15 = 66.

Superpowers 58

Harness reading of the framework checklist, the one harness2 and agent-zero used. 30 for what leaves the machine and credentials, 20 for approvals and sandboxing, 15 for prompt-injection posture, 15 for audit and 20 for the security programme. No credentials of its own and no network access in the hooks or plugins. The visual companion's logo request carries the version to primeradiant.com by default, disclosed with opt-outs, and some installs (OpenCode's git+https spec, git installs) follow main unless pinned, so new hook code runs at the next session start (22). Worktrees isolate work, reviewers are read-only, merges are confirmed, a discard needs the typed word, force pushes need the person, and issues are filed only after approval. No sandbox of its own, and plan execution runs for hours without check-ins (12). The companion has a per-session key and refuses symlinks and path escapes since 16 June 2026, the Windows hook ignores planted bash.cmd since 9 October 2026, and diagnosing-superpowers treats transcripts as evidence, not instructions. No guidance on untrusted content in repositories or fetched pages (7). Plans, briefs, reports, review diffs and a ledger are written to files, and the diagnosing skill cites path:line for each finding (11). No SECURITY.md, no security.txt and no advisories; GitHub shows a Report a vulnerability link, and both 2026 security fixes are credited in the release notes (6). 58 in all.

GitLab MCP Server 62

Hosted reading for GitLab.com. Status page at status.gitlab.com (status.io) with an API and a GitLab Duo component (20). Its incident history lives under /pages/, which the site's robots.txt disallows; our reader fetched that history page once before the rule was checked, so its contents are set aside and not scored. The front page shows only the current state, all operational on 6 November, so there is no readable history (5). MCP limits are published per plan, 60 a minute and 3,600 an hour on Free, 600 and 5,000 on Premium and Ultimate (15). 429 responses carry Retry-After and RateLimit-Reset, and a rate-limited upstream API comes back as a rate_limited tool result; accept_merge_request requires the head SHA, but we found no general idempotency guidance for writes (12). No uptime SLA on the pricing page (0). Generally available since 19.5 (10). 20 + 5 + 15 + 12 + 0 + 10 = 62.

Superpowers 68

Local-package reading, as for ai-job-search and the other skill workflows. Installs through official plugin channels (Claude Code's official marketplace, the Codex, Cursor, Grok and Kimi marketplaces) or a git install in 16 coding agents, with harness versions stated for OpenCode 2.0.4 and Claude Code's hook quirks, but bash and Node needs are spread across the docs rather than stated once (17 of 20). Still no .github/workflows directory on 2 November, so no public CI. About 60 test files in tests/ cover the hooks, plugin manifests and the visual companion server, run locally, with LLM evals in a separate repository (8 of 25). 194 issues open on 30 October 2026; 25 issues were closed between 10 and 22 October, each with a reply, while 13 of the 25 newest open issues had no comment, per the 30 October check, not re-read on 2 November (13 of 25). Dated release notes with Breaking Changes and Visible Changes sections and semver majors for them (15). Version 7.0.1 (15). 68 in all.

GitLab MCP Server 66

About 47 tools by default, plus 5 Duo tools where Duo is on, so more than 30 (5), with toolset and single-tool headers to cut the list (plus 10). Cursor pagination with first and after on most list tools, include facets one at a time, byte offsets and limits on files, jobs and artifacts (20). Documented error cases and a structured rate_limited result (15). The docs don't say whether tools carry readOnlyHint or destructiveHint; accept_merge_request has a SHA guard, and 19.4 governance sets write and delete tools to Always Ask for Duo (8). Sensible default toolsets, but no official SDKs for the REST API in two languages were found on the pages read (8). 15 + 20 + 15 + 8 + 8 = 66.

Superpowers 71

Framework reading, how much a tool-calling agent has to carry. The bootstrap (3,192 bytes) is injected at every startup, clear and compaction, and 15 skill descriptions add about 2,500 characters. Bodies load on demand and long guidance moved to references/ in v7.0.1, but single skills run to 19,506 bytes and the bootstrap asks for a skill before any reply (17 of 25). The work can be sized. Brainstorming does a quick task at once and writes a design only for a project, execution can be inline (cheapest) or one subagent per task, and every dispatch must name a model (15 of 20). Implementers report DONE, DONE_WITH_CONCERNS, BLOCKED or NEEDS_CONTEXT and the controller acts on each; review-package rejects bad ranges (14 of 20). A progress ledger and task-done records let a run resume, work happens in a worktree, merges are confirmed and a discard needs the typed word (15 of 20). Skills trigger without a command and install in one step per harness; no SDK applies (10 of 15). 71 in all.