Agents Built on Jev, Compared

12 reviewed agents compared

Jev is TypeSafe AI's System One model, released in September 2026, and these are the reviewed agents in the catalogue that use it: coding-agent harnesses, MCP servers, permission gates, browser automation and more. Most are days or weeks old, so expect small star counts, and read the FARS confidence next to each score. New arrivals are picked up automatically. If you are new to Jev itself, start with the topic guide.

New to Jev? Read the topic guide →

# Agent FARS Trust Stars License Language Last commit
1 jev-axi Judgment CLI 90 23/25 17 MIT TypeScript 2026-09-19
2 Jev Review 81 23/25 177 MIT TypeScript 2026-09-17
3 Pi Jev Semantic Router 80 19/25 19 MIT TypeScript 2026-09-18
4 Jev Guard 79 19/25 12 MIT JavaScript 2026-09-17
5 Supercov 75 16/25 70 MIT Rust 2026-09-19
6 Jev Browser 75 16/25 39 MIT JavaScript 2026-09-18
7 Pi Jev Auto Mode 73 18/25 17 MIT TypeScript 2026-09-17
8 Jev Social 71 17/25 16 MIT JavaScript 2026-09-20
9 TypeSafe Evaluate 70 16/25 128 MIT Go 2026-09-20
10 Jevbridge Decision Sidecar 62 14/25 26 MIT TypeScript 2026-09-20
11 Jev Drone 60 13/25 83 MIT Python 2026-09-17
12 Distill Terminal Coding Agent 47 10/25 678 Apache-2.0 Rust 2026-09-20
How this ranking is built

The order is the FARS total; stars only break ties, because popularity measures attention rather than whether something holds up. Two limits are worth knowing before you read the order. These are static reviews — source, documentation and repository metadata, with no runtime reproduction — which caps reliability, effectiveness and verifiability and puts the achievable maximum at 77 rather than 100. And a gap of a few points sits inside the noise of that method. Read the table as a shortlist worth investigating, not as a verdict; every score links to its full dimension breakdown and the evidence behind it. Review method →

1. jev-axi Judgment CLI

Where it differs

Unlike foreman, which runs Jev as a supervisor above a coding agent, jev-axi does not launch the agent; it connects similar completion and progress judgments to the agent's existing hooks. It also differs from generative coding models: Jev returns typed probabilistic judgments and never writes prose, so jev-axi complements rather than replaces code generation and code comprehension.

Full review → GitHub →

2. Jev Review

Full review → GitHub →

3. Pi Jev Semantic Router

Where it differs

Compared with Pi's normal compaction, Jev-guided compaction evaluates tool-history entries to retain important paths, errors, constraints, and results; it falls back to Pi's built-in summary when Jev is unavailable, unconfigured, or returns unusable data. For rapid workflow classification, agent: "jev" is positioned as a typed decision node that avoids spawning a heavyweight LLM process, rather than as a full execution subagent.

Full review → GitHub →

4. Jev Guard

Where it differs

Claude Code auto mode uses a separate classifier model to review actions. Jev Guard implements a comparable allow, ask, and deny flow through typed Jev questions such as risk, user_requested, and from_untrusted. Its distinguishing scope is the same policy across Codex, Copilot CLI, Gemini CLI, Cursor, pi, OpenCode, and ACP, while also serving as a second opinion inside Claude Code. The tradeoff is dependence on the external Jev service and weaker ask behavior on hosts whose hook APIs cannot request approval.

Full review → GitHub →

5. Supercov

Where it differs

Compared with coverage workflows that make an agent load a large HTML report, Supercov exposes small, paginated gaps queries and verifies gains through diff. Compared with hosted coverage services, its coverage execution stays under the local .supercov/ directory and does not contact a Supercov service; quality scoring is separate and does depend on TypeSafe AI's Jev API. It is also neither a test runner nor an agent platform: it wraps an existing test command and supplies quality and coverage evidence.

Full review → GitHub →

6. Jev Browser

Where it differs

The repository estimates context usage against a Playwright-MCP-style snapshot loop on the same tasks. It reports roughly five times less context per task at the median and about 8,000 versus 557,000 tokens overall, because page reading moves to Jev. Large pages account for much of that difference, while tiny pages show no saving. The source does not provide a comprehensive comparison of success rate, monetary cost, or total feature coverage.

Full review → GitHub →

7. Pi Jev Auto Mode

Where it differs

Its deterministic pattern catalogue is adapted from @nilskluewer/pi-auto-permission-gate, and the README relates its fast-path → hard-rule → classifier structure to both that extension and Qwen Code's Auto Mode. The documented distinction is the TypeSafe Jev judgment layer, combined with fail-closed outcomes, meaningful two-sided thresholds, one API request per judgment, and recorded per-condition probabilities.

Full review → GitHub →

8. Jev Social

Where it differs

Compared with calling socai directly, Jev Social adds a per-observation decision loop in which Jev selects the next operation from changing candidates and compiles cards, tables, and an evidence report. Direct socai commands are the simpler option when the platform and exact operation are already known. Jev Social explicitly does not hand browsing to another research agent or depend on socai research.

Full review → GitHub →

9. TypeSafe Evaluate

Where it differs

Compared with asking a general-purpose language model for prose and then parsing its answer, Evaluate returns typed judgments with probabilities and removes much of the prompt-format and parsing work, at the cost of depending on Jev. It offers two access routes: the TypeSafe API, which takes precedence when both credentials are set, and OpenRouter Decisions, which bills the user's OpenRouter account but still uses an alpha endpoint path.

Full review → GitHub →

10. Jevbridge Decision Sidecar

Where it differs

Compared with asking a generative LLM to produce prose and make every operational decision, Jevbridge constrains decisions to noul, choice, and score outputs and applies a separate confidence gate. Native TypeSafe Jev is the preferred backend, while Codex, Claude, Grok, OpenCode, or a generic OpenAI-compatible model can provide the System One role without a TypeSafe key. The heuristic backend works offline but is intended only for tests, not as a substitute for safety-sensitive model decisions.

Full review → GitHub →

11. Jev Drone

Where it differs

The documented alternative is the same stack with Jev disabled and a greedy “steer toward the wider side” heuristic substituted. Across three baseline runs it remained collision-free but stopped at station 2 at 17.7 m because it could not represent flying over the low beam. The reported single Jev run completed the 77.5 m course, kept the target visible 82% of the time, spent 9% of the run pinned in the reflex layer, and also had zero collisions. This is not a seed-matched statistical result: an earlier, simpler three-seed comparison found no advantage for Jev.

Full review → GitHub →

12. Distill Terminal Coding Agent

Where it differs

Compared with a terminal coding setup that sends every request to one main model, Distill can route eligible calls to a same-conversation Worker and assign bounded extraction, summarization, or compression jobs to a Utility model. Unlike an exclusively cloud-hosted setup, it can also use a local OpenAI-compatible server, though the user must operate that server and stay within its capacity and context limits.

Full review → GitHub →