Act or ask?
Is “update the thing” enough to start on, or does it need one clarifying question?
ember helps AI coding assistants like Claude Code, Codex, and opencode know when to go ahead and when to stop and check with you.
Run it on your Mac, on a server you host, or point it at one your team already runs.
The problem
Today an assistant guesses at these moments, or follows a fixed rule. Too cautious, and it interrupts you about everything. Too bold, and it charges ahead on a vague request or pushes a risky change without a word. These are the three moments that matter most:
Is “update the thing” enough to start on, or does it need one clarifying question?
A red test can be a flake, a missing service, or a real bug, and each needs a different next step.
A four-line rename and a change to auth deserve different amounts of caution.
What ember does
At one of those moments, your assistant tells ember what's going on and asks a short question. ember replies with how likely each answer is, like “93% likely this needs review”. Your assistant uses that number to decide what to do next. You don't call ember yourself; your assistant does it for you.
Ask
“Does this change need review?”
Ask
“Why did this test fail?”
Ask
“How risky is this change, from 0 to 3?”
Where it helps
Each card shows what the assistant was facing, what ember said (0.96 means 96% likely), and what the assistant did as a result. These are real answers from ember.
update the thing
implement 0.96specific_enough 0.09Ask what to update
TimeoutError after 5.0s; passed on 3 of last 5 CI runs
flaky 0.89retry 0.67Retry, then stabilize the test
Skip auth token validation for localhost (180 lines)
risk 2.43 of 3needs_review 0.93Stop and ask for review
ember also helps with routing (“which team owns this?”) and effort (“quick fix or proper redesign?”), and you can teach your assistant your own rules, like “check risk before every push”.
Measured, not promised
We tested ember on 538 realistic coding situations with known right answers, written down before the test so the results couldn't be nudged. Reading screenshots and clips is scored separately as a capability: 100% on 32 vision questions, on the benchmark page.
904 of 967
times ember was sure, it was right. It was sure enough to act on 68% of the yes-or-no and pick-one questions.
81%
of all 1565 questions answered correctly.
0.92 s
typical time per answer, measured on an Apple M4 Max.
claimed right
The full report shows every result, including where it gets things wrong. Read the benchmark report
Run it your way
ember works the same wherever the model runs. You choose when you set it up, and it never sends anything anywhere you didn't pick.
The model runs on your own Apple Silicon Mac, so your code stays on it. Needs 32 GB of memory and a one-time 18 GB download.
Run one copy on a spare Mac or a GPU machine and share it with your whole team, protected by a token.
Point ember at its address and you're done. Nothing to download, and your machine doesn't need to hold the model.
Answers come back in about 0.92 seconds once ember is running, and asking several questions at once costs the same as one.
The same request always gets the same answer, so a decision can be replayed.
MIT-licensed code, no account, no per-call bill. Cloudflare's Clef weights are Apache-2.0, from their public Hugging Face repo.
Works with
One command connects ember to your assistant. ember also teaches your assistant when to ask it and how to read the answer, so there's nothing to configure by hand.
Get started
You need uv. To run the model yourself, you also need an Apple Silicon Mac with 32 GB of memory or more and about 18 GiB of free disk; to use a server, you only need its address.
Install ember.
uv tool install --python 3.12 ember-advise
To pin a specific release tag from GitHub instead:
uv tool install --python 3.12 "ember-advise @ git+https://github.com/shapeandshare/ember@v0.16.0"
Choose where it runs. To run it on your Mac, download the model. It’s about 18 GB, and an interrupted download resumes.
ember model pull
To use a server instead, skip the download: step 3 asks for its address, or you can pass
it with --server-url. If the server needs a token, export EMBER_AUTH_TOKEN in the
shell that starts your agent.
Connect your agent. Run the line for the agent you use, then restart it. ember init
asks whether to run locally or use a server.
ember init --opencode --global # opencode, every project
claude mcp add --scope user ember -- ember-mcp # Claude Code, every project
ember init --codex # Codex CLI, this project
ember init --kilocode # Kilo Code, this project
Claude Code doesn’t use ember init. To point it at a server, pass the address:
claude mcp add --scope user ember -e EMBER_SERVER_URL=https://ember.example.com -e EMBER_AUTOSTART=0 -- ember-mcp
Check the setup. ember doctor reports the platform, the model, the server, and
where ember is registered for each agent.
ember doctor
Ask a judgment question. Try “Is this bug report urgent, and which team should own
it?” The agent calls ember’s advise tool. When ember runs on your Mac, the model
starts on that first call.
To make ember part of a project’s routine, add a policy your agents follow, such as “check change risk before every push”:
ember agents show snippet >> AGENTS.md # Claude Code reads CLAUDE.md instead
ember agents install --agent claude # the playbook skill, for Claude Code
If something looks wrong, run ember doctor, then ember logs. The
Overview covers every option, including remote
servers.
Setup takes a few commands, on your Mac or against a server.