Provides the Jev Impact dashboard and install script to visualize Prometheus-scraped impact metrics in your own Grafana instance.
Exposes Jev impact metrics from the local event log as Prometheus series via jev meter serve, allowing Prometheus scrapers to collect them.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@jev mcp triage this CI failure transcript"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP-first toolkit for TypeSafe/Jev — the System
One decision model. One stdio server (jev mcp) serves any MCP-capable
harness, backed by one local event log and Prometheus impact metrics you can
scrape into your own Grafana.
Built as an Effect codebase (v4 RC): services with Layers, typed errors,
Schema-validated boundaries, no runtime dependencies besides effect.
flowchart LR
H["MCP clients<br/>any harness"]
MCP["jev mcp<br/>typesafe_ask + tools"]
CLI["jev CLI<br/>hook · triage · check · audit · label"]
API["TypeSafe API"]
LOG[("events.jsonl<br/>local only")]
METER["jev meter"]
GRAF["Grafana · Jev Impact"]
H --> MCP --> API
CLI --> API
MCP & CLI --> LOG --> METER --> GRAFJudgment surface — typesafe_ask (MCP, jev mcp) with choice, noul,
and score primitives, next to task-shaped tools that wrap the question packs:
typesafe_verify (claims against evidence) and typesafe_review (quality
dimensions with direct before/after directions). jev ask is the same judgment
over stdin for harnesses without MCP. Details in docs/mcp.md.
Pack lab — jev eval pack replays labeled fixtures through a pack,
repeating identical calls to expose answer drift and reporting agreement,
confidence, tokens, and latency before thresholds are trusted.
Deterministic triggers — jev hook prompt prints the Jev directive when
the latest prompt matches local claim patterns, and jev triage failure
classifies failures with the loop breaker built in. The live trigger is a local
regex; the audit's detection is model-first. Details in
docs/cli.md.
Triage packs — reviewer findings (class · severity · evidence → blockers /
cosmetic / questions, with an evidence floor), harness and CI failures
(classification plus a semantic loop breaker), commit conformance (code checks
plus Jev rules, with repo spec profiles via --spec), session labels (outcome,
friction, waste, task type), and claim detection/alignment (batched per
message, batched per session).
Audit and measurement — jev audit run detects quantitative claims agents
actually made and scores compliance; jev audit prompts measures the live
trigger against real user prompts. Both are model-first: prose-only state
(fenced code stripped, credentials redacted, clipped), batched calls, code
applies only documented thresholds. Details in docs/cli.md.
Impact metrics — every call, opportunity, triage, and label lands in one
local JSONL event log; jev meter serve exposes Prometheus series for any
scraper, rendered as the Jev Impact dashboard on a Grafana instance of your
own. Details in docs/metrics.md.
Related MCP server: Jev MCP
Needs Node >= 26 and a TypeSafe API key.
git clone git@github.com:jbt95/jev-toolkit.git
cd jev-toolkit
npm ci # dev deps only: the CLI runs from source
scripts/install.sh # link ~/.local/bin/jev -> bin/jev
export TYPESAFE_API_KEY=… # must be visible to harness processesWire your harness: docs/mcp.md — any stdio MCP client, command: jev, args: ["mcp"]. Per-harness plugins (opencode, Claude Code, pi, omp) are
under integrations/ with their own install steps.
jev ask </path/to/payload.json # raw judgment over stdin
jev triage failure --transcript FILE # failure classification + loop breaker
jev audit run --since 24h --dry-run # what claims did agents make?
jev route skills --task TEXT --skills-dir DIR # pick the skill for a task
jev meter serve # Prometheus on 127.0.0.1:8788Optional: an always-on meter and a nightly label/audit run ships as launchd jobs — docs/metrics.md.
Command | Purpose |
| Raw judgment: |
| Classify a failure; transcript mode selects the failing snippet; loop breaker built in |
| Route findings: blockers / cosmetic / questions |
| Commit conformance; |
| Detect claims agents made; compliance summary; |
| Measure the live prompt trigger against real prompts |
| Outcome, friction, and waste labels per session |
| Pick the skill for a task: |
| Replay labeled fixtures through a pack; agreement, drift, tokens, and latency |
| Tail the local event log |
| Harness hook adapter (prints the directive or nothing) |
| MCP server: |
| Prometheus metrics from the event log |
Page | Covers |
System map, call lifecycle, services, event schema, audit pipeline, loop breaker, privacy boundary, design rules | |
Every command with flags, examples, outputs, exit codes, and flow diagrams | |
| |
Metric catalogue, dashboard, launchd meter, smoke test |
src/core/ services + schema (client, events, metrics, loops, text, transcript, paths)
src/mcp/server.ts stdio MCP server (jev mcp) — the judgment tool surface
src/cli/jev.ts CLI: ask | events | audit | label | check | triage | hook | mcp | meter
src/question-packs/ detection, alignment, reviewer, failure, commit, session labels, skill routing
src/eval/ pack lab: fixture replay, answer drift, and agreement reports
src/audit/ message extractors (opencode DB, claude projects, pi/omp logs)
src/replay/ review-fixture capture/score measurement aid
integrations/ harness wiring: opencode plugin · claude-code plugin · pi/omp extension
tests/ offline tests + fixtures (fake transports, temp dirs)
dashboards/ Jev Impact Grafana dashboard + install script
launchd/ always-on meter + nightly label/audit plists
scripts/ install.sh · install-dashboard.sh · check-metrics.sh · nightly.sh · smoke-*.sh
tools/oxlint/ anti-slop rule groups for the lint gate (see License)
docs/ this documentation set
skills-lock.json dev-time agent skills: source + hashes, fetched locallyoxlint (vendored anti-slop generic + Effect rules), oxfmt, tsc,
vitest. All four gates run before every commit:
npm run lint && npm run format:check && npm run typecheck && npm testTests are offline by design: fake transports, temp dirs, 127.0.0.1 only, no
module mocking.
An end-to-end smoke test (live TypeSafe API plus a real opencode session) is
operator-run and never part of npm test:
scripts/smoke-opencode.sh # agent session + operator checks
scripts/smoke-opencode.sh --no-agent # operator checks onlyOne prompt through every wired harness, asserting a fresh Jev call under each harness tag — the wiring check to run after install or config changes:
scripts/smoke-harnesses.sh # opencode, claude-code, pi, omp
scripts/smoke-harnesses.sh --only=omp,pi # subsetMIT — see LICENSE. This covers the tools/oxlint/anti-slop rule
groups as well: dev tooling loaded by oxlint.config.ts as Oxlint jsPlugins,
never part of the runtime surface (effect and @typesafe-ai/sdk are the only
runtime dependencies).
Automated flows (audit, label, triage, check, hooks) sanitize before
anything leaves the machine: credentials are masked, fenced code blocks become
[code], and long text is clipped with a marker. TYPESAFE_API_KEY is read at
call time and never logged. Raw code, diffs, transcripts, secrets, and
credentials never go into the event log; the log is local-only.
jev ask and the MCP typesafe_ask tool accept caller-provided state and send
it as provided — callers are responsible for redaction there. Details in
architecture.md.
This server cannot be deployed
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
Guarded MCP server for agent-readable business truth, provenance, readiness, and discovery.
Build and run grounded business agents over MCP: agents, knowledge bases, skills, Storylines.
