| Rating |
|---|
| AgentsWisdom score | 70 | 67 |
|---|
| Confidence | high | high |
|---|
| Rank in category | No. 2 of 7 | No. 4 of 7 |
|---|
| Pillars |
|---|
| Adoption & momentum | 76 | 67 |
|---|
| Experts (independent reviews) | 72 | 64 |
|---|
| Crowd (community sample) | 60 | 70 |
|---|
| Community posts sampled | 18 | 22 |
|---|
| Experts vs crowd | Crowd cooler by 12 | Crowd warmer by 6 |
|---|
| Adoption |
|---|
| GitHub stars | 15k | 24k |
|---|
| npm weekly downloads | n/e | 20k |
|---|
| PyPI monthly downloads | n/e | n/e |
|---|
| VS Code installs | n/e | n/e |
|---|
| Reported users | n/e | n/e |
|---|
| Reported customer organisations | n/e | n/e |
|---|
| Releases, last 90 days | 29 | 14 |
|---|
| Benchmarks (shown, not scored) |
|---|
| Terminal-Bench 2.0 | not listed | not listed |
|---|
| SWE-bench | not listed | not listed |
|---|
| Independent reviews | 6 | 8 |
|---|
| Trust & safety |
|---|
| Overall grade | A | B |
|---|
| Permission model | B | C |
|---|
| Data access scope | B | n/e |
|---|
| Data storage | A | A |
|---|
| Data retention & training | n/e | n/e |
|---|
| Incident history | A | A |
|---|
| Compliance | A | D |
|---|
| Transparency | B | A |
|---|
| Price |
|---|
| Price | Free, paid from $20 | Free |
|---|
| Plans | - Free: Free
- Pro: $20 / monthly
- Enterprise: Custom
| |
|---|
| Product |
|---|
| Category | Orchestrators & meta-harnesses | Orchestrators & meta-harnesses |
|---|
| Maker | Superset Inc. | Ping Labs (ping.gg / T3 Tools) |
|---|
| Licence | Elastic-2.0 | MIT |
|---|
| Use it via | Desktop app, CLI, MCP server, API, Mobile app | Desktop app, Web app, Mobile app, CLI |
|---|
| Runs on | macos, linux | macos, windows, linux, web |
|---|
| Models it drives | Claude Code, Codex, OpenCode, Cursor Agent, Gemini CLI, any terminal-based coding agent | Claude (via Claude Code), GPT (via Codex), Cursor models (via Cursor CLI), Grok (via Grok Build), Gemini (via Antigravity), any provider supported by OpenCode |
|---|
| Key features | - One git worktree, branch, terminal set and port range per workspace
- Runs any CLI coding agent unmodified in a persistent terminal (sessions survive app restarts via a background daemon)
- Built-in diff viewer, PR review and CI check view
- In-app browser pane with per-workspace port detection and a Design mode for UI edits
- Agent status sidebar with notifications when an agent needs attention
- Scheduled automations, CLI, TypeScript SDK and MCP server
- Remote access to workspaces on other machines and an iPhone app (Pro)
| - Drives locally installed agent CLIs: Codex, Claude Code, Cursor, Grok Build, OpenCode, Antigravity
- Separate git worktree per agent thread
- Inline diff review and one-click commit, push and pull request
- Per-thread permission modes: Supervised, Auto-accept edits, Auto, Full access
- Remote control from iOS/Android/web via T3 Connect, LAN pairing or Tailscale
- Can run as a background service (`t3 service install`)
- Switch model/provider mid-thread
|
|---|
| Verdicts |
|---|
| Praised for | - Terminal-first design lets users run the real Claude Code/Codex CLIs with their latest features, which reviewers contrasted with Conductor's custom chat UI
- Supports many agents rather than one or two vendors
- Persistent terminals that survive app updates or closing were singled out by users
| - Free and open source; uses the subscriptions you already pay for (Thalhammer, Better Stack, HN comments)
- Remote/mobile control of agents running on your own machine (Thalhammer, Geniar, HN and Reddit users)
- Worktree per agent makes parallel agents easier (HN user, Better Stack)
|
|---|
| Criticised for | - Some users reported terminal rendering glitches and freezes
- Reviewers and users found the PR/workspace flow less polished than Conductor and the UI growing 'bloated'
- Linear and some team features are gated behind the $20/user/month Pro plan; sign-in to the vendor cloud is required for the hosted build
- No Windows build; Linux build is experimental
| - Early software with rough edges and bugs, acknowledged by the project itself (Geniar, AICodeKing, HN comments)
- Default permission mode for new threads is Full access (no approval prompts)
- Missing conveniences such as queuing multiple messages (Geniar)
|
|---|
| Evidence |
|---|
| Evidence collected | 26 Sept 2026 | 26 Sept 2026 |
|---|
| Sources cited | 5 | 4 |
|---|