Cursor parity checklist: what the framework covers, what is missing, what is skipped on purpose

[suleimansh]

The direction: an open-source alternative to Cursor's agent product, built on the skills architecture. "Inferior is fine, but every important feature must be covered." This is the list to measure that against: every feature on https://cursor.com/product, one row each, against FEATURES-SPEC.md as of today.

Three verdicts. Covered: a feature in FEATURES-SPEC.md does it. Partial: part of it exists. Missing: nothing does it, with the skill or dashboard piece that would. Skipped: not ours by design, with the reason.

The table

Cursor feature	verdict	where, or what it would take
Delegate work to agents	covered	composer, presets, start from a ticket row or a queue entry
Background agents	covered	unattended runs; the queue drain; routines
Cloud agents	covered	GitHub Actions runner; Claude web through the extension; a remote machine's daemon
Parallel agents on one codebase	covered	one worktree and branch per agent; fan-out, one agent per queue entry
Multi-model, subagents	partial	model picker per run; no subagent orchestration, the queue is the answer to "nobody waits for a queued agent"
Desktop editor	skipped	editor-agnostic by design; "open in editor" on every agent
CLI	partial	the agents are CLIs; our own CLI only serves the dashboard. Missing: start an agent from a terminal
Web	covered	the dashboard; reachable from another machine behind a token
Mobile	missing	no app; a dashboard that works on a phone is the cheap version
GitHub integration	covered	issues import, PRs, auto-merge, CI watch with fix agents
Slack integration	partial	Discord notifications only
Linear integration	missing	an import preset like "Update from GitHub", over Linear's API
JetBrains integration	skipped	editor
Plan mode with clarifying questions	covered	PLAN gate with approve/decline; ranked interpretations; questions as answerable cards; ticket plans
Design mode	skipped	visual editing is the editor's
Debug mode with execution data	partial	the agent drives a real browser: console, network, DOM, screenshots. No runtime instrumentation of the user's app
Terminal command execution	covered	the CLI agent runs commands in its own checkout
Code review, BugBot	missing	a review skill: an agent reviews every PR on open and comments. Today only post-merge follow-ups and CI-watch fix agents
PR creation and management	covered	push, PR named by the agent, auto-merge armed by config and the agent's ready signal
Deep codebase understanding, fast search, large codebases	skipped	the CLI agent's job, not ours
Team rules and conventions	partial	SYSTEM.md, the-framework.yml, custom presets saved into the repo; the harness's own CLAUDE.md. Nothing says so on one page
Shared checkpoints, rollback	skipped	every agent is a branch; git is the checkpoint
@-mentions of files	covered	@ files and projects in the composer
Image as context	partial	browser screenshots inline; no upload of your own image
Plugin marketplace	missing	skills catalogues, planned
Custom skills and prompts	covered	custom presets, SYSTEM.md, the skills
MCP tools	partial	MCP packages exist in the repository; not a user-facing feature yet
Figma as a data source	skipped	design
Tab completion, inline edit	skipped	editor

Counts: 13 covered, 7 partial, 4 missing, 8 skipped on purpose.

What the page does not list and we have

• Runs on the Claude or Codex subscription the user already pays for. Cursor bills for the models itself.
• Quota-aware autonomy: unattended work stands down at the pro-rated share of the week, work a person asked for is never starved.
• Tickets, plans and claims on a branch of the repository, so agents on different machines never double-work.
• Routines: recurring unattended work, each switchable.
• Everything is git conventions and skills, so it works without the dashboard.

Proposed order for the gaps

1. Code review before merge. The largest missing item and the most visible one on the page. A skill, since it is "review this PR" wherever the agent runs.
2. Start an agent from a terminal. One command, calling the driver. The dashboard already does exactly that.
3. Team rules on one page. Nothing to build: say where the rules live.
4. MCP as a user-facing feature. The code exists; the decision is what a user sees.
5. Image as context in the composer.
6. Slack beside Discord, same notifications.
7. A phone-sized dashboard. Cheap version of mobile.

Linear and the marketplace wait for the catalogues and the distribution design.

Decisions wanted on this issue: which of the four missing items are in, and whether any skipped row should not be.

🤖 curated · Fable 5.1, effort high