User repository on GitHub
Go harness for Jev/Kev decision models: TypeSafe, Vercel AI Gateway, and in-process Kev via llama.cpp
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Agents & MCP
Gojev
A Go harness for System One decision models, covering hosted Jev through the Vercel AI Gateway or the native API plus Kev running in-process on llama.cpp, with typed helpers for gating decisions and a CLI for trying a state and questions from a shell.
- Stars
- 1
- Language
- Go
- License
- 0BSD
- Published by
- Community
- Author
- taigrr
- Last verified
- Oct 2026
About this project
Go harness for System One decision models: TypeSafe's hosted Jev, the open-weight, self-hostable Kev, and Cloudflare's Clef on Workers AI. Decision models do not generate text. They take a state (a string or structured JSON) plus a map of typed questions and return one calibrated answer per question in a single forward pass. The protocol layer is in the taigrr forks of fantasy and catwalk: fantasy.EvaluationModel / fantasy.EvaluationProvider — the modality abstraction.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Go
Highlights
- Bool, Choice of up to 255 options, and Score over 2 to 10 ordered levels, each returning the answer with a probability per option or level.
- Kev runs in-process through llama.cpp via yzma with no cgo and nothing embedded in the binary; weights are fetched and cached under $KEV_CACHE.
- Everything downloaded is verified against a digest compiled into the module, and a tampered library returns ErrLibrariesCorrupt and is never dlopen'ed.
- Classify with an Above(0.8, 0.3) band, so a decision has to be both confident and unambiguous before you let a tool call through.
Watch out
The protocol layer lives in the author's forks of fantasy and catwalk, so adopting this means taking three modules rather than one. Local Kev needs weights and the llama.cpp shared libraries fetched on first use, and WithAutoLibraries will install a pinned, verified llama.cpp if it is missing. Bumping the pinned llama.cpp version makes existing installs report ErrLibrariesOutdated until the staged, verified swap completes. The latency table comes from an M4 Max with the model already loaded, and CPU numbers are far worse, so prefer the 0.8B checkpoint without a GPU. The zero data retention flag needs a Pro or Enterprise gateway plan. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Yes. The Kev provider runs Kev entirely inside your process through llama.cpp via yzma, so there is no cgo and nothing embedded in the binary; weights and the llama.cpp shared libraries are fetched on first use and cached under $KEV_CACHE, defaulting to a kev directory in the OS cache. If you would rather keep a separate Python server, a running kev.serve is still reachable through the typesafe provider with a base URL override.
That is the part the trust model is built around. A compiled-in digest pins each checkpoint's manifest.json and every file in the bundle is checked against it, the llama.cpp release and its digest manifest are pinned the same way, and installed files are re-hashed on every load. A corrupt bundle is re-fetched, an untrusted manifest returns ErrManifestUntrusted, and a tampered library returns ErrLibrariesCorrupt without ever being dlopen'ed.
On an M4 Max with f16 weights loaded, the 0.8B checkpoint answers one short-state question in 14 ms, three questions over about 100 tokens in 61 ms and eight in 130 ms, while the 4B takes 45 ms, 205 ms and 423 ms and the 9B takes 68 ms, 323 ms and 800 ms. Multiple questions share one decode of the state through the KV prefix cache, so cost grows with branch tokens rather than state times questions. On CPU the same checks are far slower, so the README suggests the 0.8B checkpoint there.
Classify takes your state and a set of verdict options with descriptions and returns a decision you can test with an Above pair, where the first number is the confidence floor and the second is the margin over the runner-up, so a decision must be both confident and unambiguous. AskBool with an Above threshold does the same for a yes/no gate, and Fallback sends the state to a hosted model or the local Kev depending on what you pass.
Run go get github.com/taigrr/gojev, build a provider (vercel, typesafe, cloudflare or kev), get an EvaluationModel from it, then call gojev.Evaluate with a state and a map of questions. There is also a CLI via go run ./cmd/gojev or the npm launcher npx @taigrr/gojev.
Vercel AI Gateway for typesafe-ai/jev, TypeSafe's native POST /v1/systemone on api.typesafe.ai, Cloudflare Workers AI for Clef and Clef-Flash, and Kev fully in-process through llama.cpp. Clef also supports up to four embedded PNG, JPEG or WebP images as an extension to the System One API.
Bool asks yes or no and returns the probability of yes. Choice picks one of up to 255 options and returns the chosen option plus a probability per option. Score rates on an ordered rubric of 2 to 10 levels and returns an interpolated score plus a probability per level.
More Jev resources like this
16GitHub stars
Shell wrapper that intercepts unknown subcommands, reads the target CLI help output, and asks Jev which documented subcommand matches the intent behind the typo.
Jev Agents & MCP#community#jevbest#catalog-import
Community
syumai/jevyoumeanGo
syumai/jevyoumeanGo40GitHub stars
Claude Code plugin with jev find, jev ask, and a Read hook that keeps large-file context out of the agent by scoring relevance with Jev.
Jev Agents & MCP#community#jevbest#catalog-import
Community
BorisLeMeec/jevGo
BorisLeMeec/jevGo6GitHub stars
Local Windows gateway that fronts agent loops with stop, extract, and scan endpoints plus a System One-shaped typed decision path. Runs a small coder model over llama.cpp at localhost with telemetry and self-tests.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
jagsan-cyber/reflex-gateGo
jagsan-cyber/reflex-gateGo3GitHub stars
Small Go triage tool that sends each message plus three questions to Jev in one request, then prints an urgency-sorted table, a low-confidence review list, and an HTML report.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
boldbug1/jev-triageGo
boldbug1/jev-triageGo3GitHub stars
Go MCP server exposing Jev through OpenRouter or TypeSafe as five tools: decide, classify, verify, rerank and list_models. Each tool call is one paid inference request, answers are validated against the question asked, and the unchanged provider response is returned with its usage cost.
Jev Agents & MCP#community#jevbest#catalog-import
Community
freepik-company/jev-mcpGo
freepik-company/jev-mcpGo1GitHub stars
Archived Go proxy that sat between a coding-agent CLI and the upstream API to rewrite tool schemas and history with Jev, kept for the negative result: tool steering cut tokens on one model and added 37% and 51% on two others.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
nekowasabi/jev-routingGo
nekowasabi/jev-routingGoFrom the Jev community
Posts from builders shipping with Jev right now.