OpenInterpretability
  • Ekbasis
  • Guide
  • Research
  • Lab
  • Tools
  • Notes
  • Registry
  • Manifesto
Download the modelDownload
OpenInterpretability

Ekbasis: know what an action will do before your agent runs it. From an independent lab for AI-agent safety. Open weights, Apache-2.0: download it and run it yourself.

Download the model
Ekbasis
  • Getting started
  • Download (Hugging Face)
  • For agents (agents.md)
  • Results and models
  • Client and CLI
  • Cookbook
The lab
  • About the lab
  • Open tools
  • Observatory
  • ProbeBench
  • InterpScore
  • Academy
Research
  • Manifesto
  • Roadmap
  • Papers & posts
  • Docs
Community
  • GitHub org
  • Notebooks
  • SDK (openinterp)
  • HuggingFace
© 2026 OpenInterpretability — Apache-2.0 for code, CC-BY 4.0 for docs.Built in public.
The hosted API was discontinued on 2026-10-20 · Ekbasis is now a download

Know what an action will do before your agent runs it.

Ekbasis is a consequence model. Send the state and the action; get a calibrated probability for what happens — lost work, a failed payment, a leaked file — in one call.

Download the model How to use it (5 min)

Open weights, Apache-2.0 · INT4 for vLLM, 4-bit MLX for Apple silicon · runs on your own GPU or Mac

State you send

Working tree: app.py, utils.py and README.md have uncommitted changes (2 hours of edits, never committed or stashed).

About to: git reset --hard

Question

Will uncommitted work be lost?

Ekbasis answers

uncommitted changes are gone98.5%
nothing is lost1.5%

One forward pass · 135 input tokens ≈ $0.000005 · 0.19 s on the server

Real responses of Ekbasis-27B on the hosted API, recorded on 10 Oct 2026. Nothing is generated: the answer is read from the model in one pass.

Measured, pre-registered, failures published too

Studies run by us and one third-party benchmark: every plan was hashed before the runs, and the hypotheses that failed are published next to the ones that held. Read them before you rely on the numbers.

6/24 → 1/24

tasks where Claude Sonnet did harm on real apps (Gitea, Nextcloud, Roundcube): 25.0% → 4.2%

secondary contrast, one run per task; for Claude Haiku the pre-registered hypotheses failed

Read the paper, failures included →

24/36 → 2/36

harmful runs by Claude Haiku operating a real Kubernetes cluster: 66.7% → 5.6%

asking the human every time also reached 0%, with 3× the questions; matching a Sonnet guard failed by one case

Read the study, failures included →

~50% → 0–6%

harmful runs by an open agent (GLM-5.3-Flash) in each of 3 runs of the same 47 tasks on our demo apps: 23, 24, 26 → 3, 0, 0 — two runs through this API

the same tasks and agent each time; apps built for the study, consequences hidden from the agent

Read the study, failures included →

97.5 F1

on AgentDojo prompt injections in TS-Bench, a third-party benchmark (GPT-4o published: 56.6); rewriting the injections changed none of 301 decisions

weak on its AgentHarm and ASB subsets (F1 54.6 and 64.3); our untested reading: those labels follow the request or the tool, not the step

Read the benchmark, failures included →

How it works

Not a chatbot and not a judge. A forecaster: it tells you what will happen, and your code — or a person — decides what to do about it.

1

Describe the state

The facts the action touches: files, balances, branches, recipients, rules. End with “About to: …”. The CLI writes it for you for git.

2

Ask typed questions

Choice among outcomes you list, yes/no with criteria, or a count. Several questions about one state go in one request.

3

Act on calibrated odds

Each option gets a probability. High risk: stop and ask the human. Low confidence: a fact is missing — go read it. Can’t foresee: treat it as risky.

Plug it in where your agent acts

One local server for all of them. Start with the guard; go to the API when the action is not a command.

Git and shell guard

One command. The CLI reads your repository and writes the state for you.

pip install ekbasis
export EKBASIS_URL=http://127.0.0.1:8000   # your own server: see "Run it yourself"

ekbasis git-check -- "git reset --hard"
# Ekbasis: RISKY (lose uncommitted work: 99%)  → exit 2

Claude Code hook

Before every Bash command, the hook asks you only when something could be lost.

// ~/.claude/settings.json
{"hooks": {"PreToolUse": [{"matcher": "Bash",
  "hooks": [{"type": "command",
             "command": "ekbasis-claude-hook", "timeout": 30}]}]}}

MCP server

check_git_commands, predict_consequences and preflight_command for any MCP client.

claude mcp add --scope user ekbasis \
  -e EKBASIS_URL=http://127.0.0.1:8000 -- ekbasis-mcp

Migration guard (GitHub Action)

On every pull request: will its PostgreSQL migrations fail or lose data? One comment, before the merge. How it works and what we measured →

- uses: OpenInterpretability/ekbasis/actions/migration-guard@v0.1.12
  with:
    api-url: ${{ secrets.EKBASIS_URL }}   # your own server, reachable from the runner
# Risky: …/05_add_visit_id — may fail on existing data (96%);
#        nothing in the file would be applied

REST API

Any action, any language: describe the state, ask typed questions.

curl http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"state": "Balance: $1,020.00. About to: wire $1,000 + $25 fee.",
       "questions": {"q": {"type": "choice",
         "instructions": "What happens with the wire?",
         "options": ["succeeds", "fails: exceeds the balance"]}}}'

Building an agent? Point it at openinterp.org/ekbasis/agents.md — the same guide written for language models.

The mistakes it is built to see

Git

reset, checkout, clean, branch -D, stash drop, force-push over commits the remote holds

Files and shell

rm and redirects over data without a backup, unset variables in paths, symlinks

Databases

DELETE or UPDATE without a narrow WHERE, DROP, migrations on real data

Payments and apps

transfers and fees, refunds in flight, cancellations that take the return flight too

Messages and sharing

reply-all with an external recipient, links that expose more than the folder you meant

Cloud and infra

the wrong kubectl context, compose down -v on a volume with data, overwrites on S3

Measured in 21 domains — 168 of 168 scored scenarios right, written by us with the deciding facts in the state: every result · every example with its real answer: the guide.

Run it yourself

The weights and the System One server are public. Four commands from the model card take you from download to the first check.

Hosted API discontinued · 2026-10-20

The hosted Ekbasis API no longer takes requests, and API keys and credits are no longer sold. Ekbasis is now download-only: the weights and the System One server are public on Hugging Face, and you run them yourself.

  • caiovicentino1/Ekbasis-Skopos-27B-INT4 — the model the hosted API served last (GPTQ W4A16, vLLM)
  • caiovicentino1/Ekbasis-27B-INT4 — the earlier release (GPTQ W4A16, vLLM)
  • caiovicentino1/Ekbasis-27B-MLX-4bit — for Apple silicon (MLX)
hf download caiovicentino1/Ekbasis-Skopos-27B-INT4 --local-dir Ekbasis-Skopos-27B-INT4
bash Ekbasis-Skopos-27B-INT4/serve_vllm.sh Ekbasis-Skopos-27B-INT4 8001
python Ekbasis-Skopos-27B-INT4/serve.py --model Ekbasis-Skopos-27B-INT4 --vllm-url http://127.0.0.1:8001 --port 8000
pip install ekbasis && EKBASIS_URL=http://127.0.0.1:8000 ekbasis git-check -- "git reset --hard"
Download the model

Built to be checked

Open weights

Apache-2.0, on Hugging Face: INT4 for vLLM and 4-bit MLX for Apple silicon.

Pre-registered

Plans hashed before the runs. The studies that missed their bars are published next to the ones that held.

Fails closed

No answer in time: “cannot foresee”, which every tool treats as risky.

Your data stays with you

You run the model and the server: states and questions never leave your machine.

A warning layer that can be wrong, not a security boundary. It is strongest where the state states the facts and weakest on domains far from its training — read the limits.

From OpenInterpretability

An independent lab for AI-agent safety. Ekbasis came out of our research on why agents fail; the papers, datasets and the other open tools are all public.

ResearchThe lab

Let your agent look before it leaps.

Download the model Results and models