How to use Jev with Claude Code, Codex and Cursor

Two ways Jev fits into a coding agent

A Jev + Claude Code setup has two layers, and so does one in Codex or Cursor. At build time, the agent writes your application's Jev integration: the request, the questions, the thresholds. That is what TypeSafe documents, and what its skill is for. At run time, the agent itself asks Jev during its loop — whether a patch is safe to apply, whether a claim matches the test output, whether a command can run without a prompt. TypeSafe publishes nothing for that layer; the MCP servers, hooks and plugins below are community projects. Most setups want both.

Install the TypeSafe skill in Claude Code, Codex, Cursor and OpenCode

The skill is typesafe-ai/skills (MIT, version 0.6.2 when we installed it). In Claude Code it is a plugin:

claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai --scope project

--scope project only records the plugin in .claude/settings.json; each teammate still runs the install once. For the other agents, one command writes the skill to .agents/skills, which Codex, Cursor and OpenCode all read, and links it into .claude/skills:

npx skills add typesafe-ai/skills --skill typesafe-ai -a claude-code -a codex -a cursor -a opencode -y

We confirmed Codex picks it up offline: its prompt preview listed the skill without a model call.

Add a Jev MCP server, tool by tool

We tested jkudish/jev-mcp 0.14.7, which exposes twelve tools such as jev_decide, jev_review and jev_gate and calls Jev with your TypeSafe key. Every tool can read that key from your environment, so nothing secret goes in a file you commit:

Jev + Claude Code

claude mcp add-json --scope project jev \
  '{"type":"stdio","command":"npx","args":["-y","@jkudish/jev-mcp@0.14.7"],"env":{"TYPESAFE_API_KEY":"${TYPESAFE_API_KEY}"}}'

The written .mcp.json keeps the variable reference as written, and Claude Code expands it at startup; GitHub Copilot reads the same file. Project servers sit at “Pending approval” until you run claude in the folder and accept them. The tools then appear as mcp__jev__jev_decide and so on.

Jev + Codex

# .codex/config.toml
[mcp_servers.jev]
command = "npx"
args = ["-y", "@jkudish/jev-mcp@0.14.7"]
env_vars = ["TYPESAFE_API_KEY"]
startup_timeout_sec = 30

Codex reads a project's .codex/config.toml only once you trust the folder; before that, codex mcp list said no servers were configured. env_vars forwards the key from your shell, where codex mcp add --env would write it into the file.

Jev + Cursor and OpenCode

// .cursor/mcp.json
{ "mcpServers": { "jev": { "command": "npx", "args": ["-y", "@jkudish/jev-mcp@0.14.7"],
  "env": { "TYPESAFE_API_KEY": "${env:TYPESAFE_API_KEY}" } } } }

// opencode.json
{ "mcp": { "jev": { "type": "local", "command": ["npx", "-y", "@jkudish/jev-mcp@0.14.7"],
  "environment": { "TYPESAFE_API_KEY": "{env:TYPESAFE_API_KEY}" }, "timeout": 30000 } } }

Set OpenCode's timeout yourself: its schema and its docs give different defaults. No TypeSafe key? This site's Jev MCP server is hosted and takes a Jagent key instead, with the client setup on its page.

TypeSafe also serves an MCP server for its documentation at https://docs.typesafe.ai/mcp, with search and read tools and no key. It has a third tool, submit_feedback, that sends a report to TypeSafe; deny it so an agent does not file one unasked.

What Jev checks cost inside a coding loop

The checks a Jev + Claude Code setup makes most, run forty times on 2026-11-04; all but one call was answered by jev-1.14.0, most through the jev-latest alias. Cost is input tokens at $0.042 per million; output is not billed.

The diff review is the pattern worth copying. One request asked five yes/no questions per diff, and code accepted a patch only if it fixed the task at 0.7 or more with every risk at 0.3 or less. That let through the good fix and nothing else. The patch that deleted the failing assertion scored 0.98 on “tests passed” — the remaining test did pass — so a tests-pass check on its own would have approved it.

Two results argue for measuring before you trust a default. The MCP server's jev_review separated good patches from bad ones by probability (0.78–0.80 against 0.04–0.07) yet returned escalate for all of them at its default thresholds, so read the numbers rather than the action. And a permission gate asking whether a command was undoable and stayed inside the project approved none of four safe commands at 0.9. Split into separate questions, “undoable” alone separated safe from risky (0.79–0.98 against 0.03–0.09). Box found the button in a clean screenshot, but on one with a cookie banner over it Box still answered; a companion question, whether the button was fully visible, is what caught it (0.02).

Hooks: where a Jev gate goes in each agent

• Claude Code: PermissionRequest, which runs only when Claude Code is about to ask you. Print an allow decision, or nothing to leave the prompt as it is.
• Codex: PermissionRequest in .codex/hooks.json, same contract, loaded only in a trusted project after you approve the hook.
• Cursor: beforeShellExecution, which runs before every command. A crash or timeout allows the command unless you set failClosed.
• OpenCode: a plugin in .opencode/plugins/.

Check a static list of dangerous commands before calling Jev at all: ours caught a force-push and an rm -rf in about 70 ms with no network call. Pin jev-1.14.0 once a threshold is tuned, because jev-latest moved from 1.13 in late October; the Jev versions are listed on the model page.

Setup problems we hit

• In our Jev + Claude Code setup, the server showed as connected with the key unset; the first real call then failed.
• Starting the server through npx took 2.6 to 10 seconds, against a 10-second default startup timeout in Codex. Raise it, or install the package and start it in under 0.3 s.
• Hermes passes MCP servers only a short list of environment variables, so a key exported in your shell never reaches the server unless the config names it.
• An MCP server shipped inside a Claude Code plugin names its tools mcp__plugin_…, so a hook matched on the bare server name never fires for them.

Community tools for Jev in coding agents

Stars read from GitHub on 2026-11-04; all MIT. One caution from a compaction port's own README: pruning an agent's history can drop hidden reasoning and force costly prompt-cache rewrites, so measure the bill before and after. More projects, with screenshots, are in the Jev showcase.