Documentation

Erevos-03

Erevos-03 is a coding agent that runs in your terminal through the skelmini CLI. This guide covers installation, everyday use, security work, the command reference, and billing.

Quick start Features Security Commands CLI flags Account & API key API access Pricing About Support

Quick start

Install the CLI, then sign in from your terminal. The npm install requires Node.js 18 or later.

npm
$ npm install -g skelmini
macOS · Linux
$ curl -fsSL https://skeldyne.com/install.sh | sh
Windows · PowerShell
> irm https://skeldyne.com/install.ps1 | iex
Sign in
$ skelmini
> /login

A browser window opens to confirm the sign-in. Once approved, the CLI is linked to your account — there is no key to copy or paste.

Run your first task
$ cd my-project
$ skelmini
> add a /health endpoint with a test

Features

Single model
One capable model for every task — no model picking or routing required.
Knowledge layer
Curated, continuously updated engineering knowledge — current package information, verified error fixes, and technique guides — applied automatically when relevant.
Vision
Show it what you see — paste a screenshot or point at an image file, and the agent reads charts, UI states, diagrams, and error screens directly.
Verify loop
Edits are checked against your project's own tests, with automatic re-diagnosis and retry on failure.
Plan mode
The agent explores read-only, presents a plan, and waits for your approval before modifying code.
Auto-compaction
Long sessions are summarized automatically, so work continues past the context limit without losing state.
Orchestrator
Delegate parallel work to sub-agents, including an always-on read-only explorer.
Background tasks
Persistent shells, reminders, and scheduled jobs, managed per session.
Rewind
Restore code and conversation to any earlier point in the session.
Durable plans
The task plan persists across compaction, resume, and disconnects.
Danger guard
Destructive commands require explicit confirmation, even with auto-approval enabled.
User extensions
Define your own slash commands and shell hooks; reload them live with /hooks reload.
Security review
On-demand, read-only vulnerability audit — your code never leaves your machine.

Security

There is no separate security mode. The same agent, the same plan gate and the same read-only explorer do this work — what changes is how you scope the task. Used well, Erevos-03 is good at the parts of security work that are really careful reading: tracing whether a check is actually reached, comparing a route against its siblings, and noticing the difference between code that looks safe and code that is safe. It is a tool for engineers working on systems they control. It does not replace a professional audit, and it is not a substitute for your own threat model.

Review a change before it merges

Point it at a diff, not at a repository. A diff-scoped review answers a question narrow enough to answer well: does this change introduce a defect? It reads every line of the change, checks each one against the code around it, and ranks what it finds.

> security review of the auth changes in #482 before we merge
> review this diff for injection and access-control problems, ranked by severity

A review that reports nothing is a result, not a failure — it means the change introduced no defect the agent could see. Asking about a whole file or a whole service is a different and much larger question, and the answer will be correspondingly vaguer. Narrow the scope and the findings get sharper.

Test a system you are authorized to test

Start in plan mode. The agent explores read-only, states the scope it intends to work within, and waits — so you see the plan before anything runs, and you can correct the scope rather than the damage.

> /planmode
> authorized test on staging: access control on the invoice routes, read paths only

Keep it on a staging target, give it two throwaway accounts rather than real ones, and let it exhaust read paths before you approve anything that writes. Ask it to reproduce a finding twice before you believe it — a result that reproduces is evidence, and a result that does not is a story.

Only against systems you own or have written permission to test. Testing a third party without authorization is unlawful in most jurisdictions, and the agent will decline to help with it. Probing Skeldyne’s own service is separately prohibited by the Terms.

The controls that keep the blast radius small
Plan mode
While a plan is unapproved the turn is genuinely read-only: write tools are denied, shell commands that would mutate anything are denied, and a long-lived background shell is denied outright. Approval reopens them mid-turn.
Read-only explorer
A sub-agent that can look but never touch, so reconnaissance costs you nothing in risk.
Danger guard
Destructive commands prompt for confirmation even when auto-approval is on.
Rewind
Restore code and conversation to any earlier point in the session, so an experiment is always reversible.
Verify loop
A fix is not finished until your project’s own tests pass — which is also how you find out that a security fix broke something else.
Your code and your credentials

The agent runs on your machine. Commands execute in your shell, in your working directory, with your environment and your credentials — which means it can use them without you ever handing them to us.

The trade-off is worth stating plainly: anything the agent reads becomes model context. If you ask it to open a file of secrets, that file is sent with the request. Keep credentials out of the working tree, and use plan mode when you are not sure what a task will open.

The semantic index is built from the files git tracks, so anything ignored by git is never indexed. That is a useful default, not a secrets guarantee — the agent can still open a path you point it at directly.

Connections are encrypted, and your API key can be rotated or reset from your dashboard at any time. Your code and prompts are not stored on our servers; what we do keep, and for how long, is set out in the Privacy Policy.

Command reference

Session
/loginSign in from the terminal via a browser confirmation.
/logoutSign out and remove saved credentials.
/newStart a fresh session with a clean context.
/clearClear the screen and scrollback, then start a fresh session.
/idPrint the current session ID.
/resumeOpen the saved-session picker — resume any prior session.
/exitQuit skelmini (Ctrl-C also works).
Model & behaviour
/modelOpen the model picker./model <name>
/effortSet reasoning effort: low · medium · high · xhigh · max./effort max
/bypassToggle auto-approval of tool actions.
Control & safety
/planmodeExplore read-only and require plan approval before any edit.on | off
/verifyConfigure the post-edit verification loop.on|off auto test <cmd> cycles <n> timeout <s>
/compactSummarize the session to free context space, or tune the automatic trigger.auto on|off|<tokens>
/rewindRestore code and conversation to an earlier turn.
/security-reviewRun a read-only vulnerability audit.[path] | git | focus
Knowledge & memory
/planShow the durable task plan.clear
/lessonsReview the techniques the agent has learned from your sessions.del <n> · clear · migrate
/memoryManage project facts kept in context — add one by typing # fact.add · edit <n> · del <n> · clear
/codesearchSemantic code search — find code by meaning. On by default; run index once per repo. The index is built and stored on your machine.index · on · off
Multi-agent
/agentsManage custom sub-agents — add, edit, or delete.
/skillsOrchestrator and deep-research workflows.
@agentSend a task to a named sub-agent.@<name> <task>
Tooling & inspection
/usageShow balance, month-to-date spend, and token usage.
/tcallsShow the session's tool-call history, with filters.
/mcpManage MCP server connections — list, test, add, remove.
/tasksView and manage background shells, workers, reminders, and agents.
/hooksManage your custom shell hooks and slash commands.on | off | reload
/remindSet a session reminder — fires once when the time elapses./remind 30s check deploy
/pastePaste text or an image from the clipboard into your prompt — images are attached as extracted text. Ctrl-V also works.
Help
/helpList every command, with descriptions, inside the app.

CLI flags

Flags available when launching from the shell. Run skelmini --help for the full list.

-p "…"Run a single task non-interactively, print the result, and exit.
--continueResume the most recent session, optionally with -p "…".
--resume <id>Resume a specific session by ID.
--model <name>Choose the starting model (default: erevos-03).
--effort <level>Set reasoning effort: low · medium · high · xhigh · max.
--yoloStart with auto-approval already enabled.
--bannerPrint the welcome panel and exit.
--modelsList available models and exit.
--versionPrint the installed version.
sessionsList saved sessions (no dash prefix).

Account & API key

Your dashboard shows your credit balance, spend for the current month, and daily usage.

Your API key

An API key is issued with your first top-up and displayed once, on the dashboard. After that it appears only in masked form. In normal use you never handle it directly — /login links the CLI to your account automatically. To call Erevos-03 from outside skelmini, see API access.

Resetting the key

If a key is lost or may have been exposed, use Reset API key on the dashboard. A new key is issued and shown once; the old key stops working immediately. Your balance and usage history are unaffected. Afterwards, run /login again in skelmini.

API access

Your Skeldyne key also works outside skelmini. The API is OpenAI-compatible, so any editor, tool, or agent framework that accepts a custom endpoint can use Erevos-03 directly — same credits, same rates, no separate signup.

Base URLhttps://api.skeldyne.com/v1
Modelerevos-03
Auth headerAuthorization: Bearer <your API key>
curl
curl https://api.skeldyne.com/v1/chat/completions \ -H "Authorization: Bearer YOUR_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "erevos-03", "messages": [{ "role": "user", "content": "Explain this stack trace: ..." }] }'
Any OpenAI-compatible SDK
# Python — pip install openai from openai import OpenAI client = OpenAI(base_url="https://api.skeldyne.com/v1", api_key="YOUR_API_KEY") r = client.chat.completions.create( model="erevos-03", messages=[{"role": "user", "content": "Write a regex that matches ISO dates"}], ) print(r.choices[0].message.content)

Streaming works the standard way ("stream": true). Usage is billed per token at the same rates as skelmini and shows up on your dashboard like any other usage.

Best results in skelmini. Erevos-03 is tuned end-to-end for the skelmini agent. Its knowledge layer travels with the API, but the agent behaviours — the verify loop, plan mode, code search, orchestration — live in the CLI. Over the raw API it behaves as a strong general coding model; inside skelmini it does its best work.

Pricing

$5minimum top-up

Pay-as-you-go prepaid credits — one flat rate at all hours, and credits never expire. Heavy users can subscribe to Pro for half-price usage.

Per 1M tokensPay-as-you-goPro
Input (cache-miss)$2.18$1.09
Input (cache-hit)$0.018$0.009
Output$4.35$2.18
Context caching is automatic. Repeated prompt content is billed at the cache-hit rate — in practice around 75% of input tokens. A typical coding task (4–5 turns, ~25K tokens) costs a few cents; a $5 top-up covers roughly 80–120 such tasks.
Pro subscription

$15/month — half-price usage at the Pro rates above, with $5 of usage credit included every month. It pays for itself above roughly $30/month of pay-as-you-go usage. To subscribe: sign in to your dashboard and click Subscribe to Pro. Cancel any time — your rate simply returns to standard and your credits stay.

ModelErevos-03
BillingPrepaid credits · optional Pro subscription
Minimum top-up$5
RateFlat, 24/7
Credit expiryNone
RefundsUnused credits, within 14 days

Balance and usage are always visible on your dashboard.

About Erevos-03

Erevos-03 is Skeldyne's coding agent — our own stack of retrieval, tooling, and tuning, built on leading open-weight foundation models (including DeepSeek, Qwen and Kimi). As we add capabilities like vision, the models underneath evolve; the Erevos-03 experience stays the same.

Support

For questions or issues, contact support@skeldyne.com from your account email address.

Legal: Terms of Service · Privacy Policy · Refund Policy