Leaderboard / #1 of 13 / Driven end-to-end

Orca

Teams and solo builders on any OS who want a GUI they can also script.

Weighted feature assessment93.0range 93.0 to 93.0doc coverage 100% / Stub end-to-endmeasured: Stub run x3, launch probe

Maker: Stably AI (stablyai)
License: MIT
Pricing: Free
Stars: 97922
Open issues+PRs: 8610
Latest release: v1.4.230 2026-10-19

Scores by category

Rubric breakdown

Read the scores as a feature assessment. Orca was run end-to-end with stub agents; T3 Code was not run through an equivalent session, so Orca 93 vs T3 Code 90 is a difference in assessed features and evidence, not an experimentally proven gap.

AAgent breadthweight 155
BIsolation and parallelismweight 155
CReview, merge and PR flowweight 154
DRemote, cloud and mobile reachweight 105
EOpenness and licenseweight 105
FMaturity and momentumweight 154
GLinux setup friction (measured)weight 105
HPricing and BYO modelweight 55
IDocs and transparencyweight 54

Only GUI orchestrator we could drive end-to-end on Linux: headless runtime, scriptable CLI, 18/18 stub tasks verified across 3 logged clean runs, survived restarts.

Verdict

Strengths and weaknesses

Strengths

Headless runtime and full CLI. We drove the whole suite without a display
Concurrent worktree creation, crash isolation, terminals that survive restarts
35+ agents, SSH remotes, mobile companion, MIT

Weaknesses

About 1.1 GB peak RSS for 6 tiny sessions
Ships several builds a day and its orchestration layer is still experimental
8.6k open issues+PRs, and a star count we cannot explain

Evidence

What we saw and where it came from

Hands-on (this box): Fully exercised headless: 6 concurrent worktrees + stub agents, 3 logged clean runs, 6/6 verified each; crash isolation; restart persistence.
Launch probe: First window 4.82 s, process-tree RSS 920 MB on Xvfb.
Lane A, 3 runs: All 6 sessions launched in 1.52 to 1.90 s; 18/18 verified; peak RSS 1124 to 1140 MB; orca serve ready in 2.79 to 3.06 s.
Crash and merge: SIGKILL on one session: the other 5 stayed verified and partial work was preserved. Git merges: 5/6 clean, with t5b conflicting in shelf/config.py as designed.
Known issues: 8.6k open issues+PRs; ships multiple builds per day (v1.4.230), so behaviour moves fast. 'Orchestration' layer is flagged Experimental in docs.
Notes: GitHub reports ~97.9k stars, unusually high for a 7-month-old repo; we did not audit star provenance.

Screenshot of Orca first window on Linux (Xvfb launch probe) — First window, Xvfb launch probe

Sources

Feature facts

From docs, repos and our probes

Platforms: macOS, Windows, Linux (AppImage/deb/rpm), iOS + Android companion; headless `orca serve` runtime
Supported agents: Any terminal agent; README lists 35+ incl. Claude Code, Codex, Grok, Cursor, Copilot, OpenCode, Amp, Antigravity, Pi, Devin, Goose, Cline, Droid, Kiro, Qwen Code...
Parallelism: Worktree per task; fan one prompt across N agents; experimental Runs/Tasks/supervised workers (DAG) via CLI
Isolation: Git worktrees; SSH remote worktrees
Review, merge, PR: Diff review; GitHub/GitLab/Linear/Jira linking; PR/MR fields on worktrees
Remote, cloud, mobile: Strong: mobile companion, remote runtimes via pairing, SSH
MCP: Not verified (CLI + bundled agent skills are the documented control surface)
Cost visibility: Account/usage features exist (e.g. 'codex reset credit', 'Configure usage tracking' in UI); depth not verified
Model choice: --model / --effort for Claude, Codex, Cursor launches
Setup: Install app (onboarding detected OpenCode and offered 44 more agents) or `orca serve`

"unknown" means we found no source and did not guess.