AmbitGet startedFAQReferenceGlossaryAll docsGitHubLive demo

Ambit and Jev

TypeSafe's Jev is a model that answers typed questions with calibrated probabilities instead of text: pick one of these options, place this on a scale, yes or no. It is fast and cheap enough to ask on every tool call, and agent harnesses now use it to decide what to try and whether a call looks safe. Ambit answers a different question: what this setup can do, and what an agent may do with it. Used together, Jev decides what to try and ambit_can decides whether it may run. Ambit also puts Jev on the map, checks that a local clone answers, and can route a goal through one.

Jev on the map

Ambit models Jev as Typed Judgment, in the Model Access era of the curated tree. An era is one of the map's columns, ordered by what depends on what. Typed Judgment is reached however Jev arrives in your config:

A clone running on your own hardware also reaches Local Typed Judgment, which sits beside Local Embeddings in the Sovereignty era. That era holds what runs without a hosted service. The difference matters. The hosted API may keep requests for a while unless your account has a zero-retention agreement, and a local clone keeps the state it judges on your machine.

Every node of the curated tree is a combo, a capability composed from others, so its id starts with combo:. ambit impact combo:typed-judgment shows what depends on it, and ambit goal typed-judgment shows what it would take to reach it.

Proving a local model answers

ambit verify combo:local-typed-judgment asks a local clone one trivial question and records whether judgments came back. It tries the ports Kev and OpenJev use by default, 8009 and 8080 on 127.0.0.1, and TYPESAFE_BASE_URL or JEV_BASE_URL when either names this machine.

It never contacts another host. A base URL that is not loopback is skipped, and one carrying credentials is refused, because http://127.0.0.1:1@host is a request to host. Typed Judgment itself has no check: proving the hosted API answers would spend your key on a call you did not type.

Routing a goal through it

ambit goal routes a sentence by the words the curated tree authors for each capability. A goal that names none of them had no route in. With a clone running locally, --judge closes that gap:

ambit goal "triage incoming bug reports" --judge

When the vocabulary cannot recommend anything, the sentence goes to the local model as one Choice question over every node in the tree, plus one option saying none of them fits, and the likeliest node comes back with its probability. Below even odds, nothing is suggested and the three likeliest are listed; when "none of these" is the likeliest answer, nothing is suggested at all. Without that option a Choice has to pick a node, and does so with confidence: Clef-flash put juggle flaming torches in Skill Library at 0.69. With it, Clef-flash, served locally from its open weights on an M1 Max, placed 44 of 48 sentences written for the tree first and all 48 in its top three, answered "none of these" for 13 of 16 that meant nothing in the tree, and kept the other three under even odds. Each answer took about six and a half seconds there. It writes nothing to the graph, and a goal the tree's words already cover opens no socket at all.

The judge must run on this machine. --judge=<url> or AMBIT_JUDGE_URL can point it elsewhere on loopback, and it defaults to Kev on http://127.0.0.1:8009. Any other host is refused before anything is sent. A server for the open Clef weights is a judge like any clone once it answers on /v1/systemone; Clef on Workers AI is a hosted API, and is refused for the same reason the hosted Jev is.

When no judge answers at all, the sentence goes to an embedding model on this machine instead: Ollama's /api/embed on http://127.0.0.1:11434, or AMBIT_EMBED_URL, held to the same loopback rule. A server that answers Ollama's path with a 404 is asked in the OpenAI shape, /v1/embeddings, which is what llama-server and LM Studio serve; a new embedding model often reaches llama.cpp before Ollama can run it. It picks EmbeddingGemma 2 if the server lists one, then any model whose name says it embeds, or the one AMBIT_EMBED_MODEL names. What comes back is the three nodes nearest the sentence in meaning, each with a cosine similarity. A similarity is not a probability, so the answer says which kind it is. The nearest node is suggested only above a floor measured for that model on sentences written for the tree, and only EmbeddingGemma 2 has one so far: on 48 such sentences its nearest node was right for 34, and every sentence meaning nothing in the tree scored below the floor. Any other model lists the nearest and suggests nothing. A judge that answered with an error is reported as one and never passed over.

Where Jev stops

Use Jev to decide what to try, and ambit_can to decide whether it may run.

Jev returns a probability, and text inside the state it reads can move that probability. In one published test, injected text dropped its probability of blocking rm -rf ~/.ssh from 0.76 to 0.48. A runtime may consult Jev before it asks Ambit, and that is a sensible use. What Jev never does is stand in for a grant. Ambit's answer to whether an action may run comes from a grant a person set in advance, and nothing an agent reads can widen it. The deep dive states the rule, and the FAQ has the short version.

See what your own agents can reach

Ambit reads the MCP configs of Claude Code, Cursor, OpenCode and thirteen more agents into one local map: every server, which ones work, what one more step would open, and what breaks if a piece goes.

Open the live demoa sample setup in your browser, nothing to install

On your machine: npx ambit-cli, with Node 22.18 or newer on macOS or Linux.

MIT licensed · one local SQLite file, no telemetry · on npm and in the official MCP registry