Essarion Build v0.4.0 · 2026
Live on PyPI Python 3.11+ Apache 2.0

Reasoning, amplified.

Point it at any model and it works like a careful engineer: it plans the change, writes it, then checks its own code before you see it. Run it as the essarion command, or import it as a Python library — your key, your model.

One line. It installs the essarion command and gets it onto your PATH — open a new terminal and type essarion.

Rather not pipe a script to your shell? pipx install essarion-build && pipx ensurepath does the same. And if the command ever goes missing, python -m essarion_build still works.

It's a normal Python package — install it and import what you need. Nothing to add to your PATH.

from essarion_build import reason, generate — there's a working example further down.

Cloud → PyPI ↗ GitHub ↗ Docs ↗
The loop

Plan, draft, defend.

Every task runs through three passes: it plans the change, drafts it, then checks its own work. You get the reasoning next to the code, so it's easy to see how it got there.

analyzingcontext · skills · constraints
thinkingtradeoffs · failure modes
planningapproach · steps · verdict
defendedcode · justification
Quick start

Two functions. One key.

Set OPENROUTER_API_KEY or ANTHROPIC_API_KEY, give it some context, and call. Two functions cover most of what you'll do.

# pip install essarion-build
from essarion_build import Context, reason, generate

ctx = Context().with_all_skills().add_repo("./")

# Heavyweight reasoning — plan, tradeoffs, verdict.
r = reason("Design a JWT parser with claim validation", context=ctx)
print(r.plan, r.tradeoffs, r.verdict)

# Reasoned generation — code with a defense paragraph.
g = generate("Implement the parser above", context=ctx)
print(g.code)
print(g.defense)
What's inside

Production guardrails, not toy demos.

54 skills

Bundled coding knowledge

API design, auth security, concurrency, debugging, refactoring, secure coding, testing, and 47 more — composable per call.

6 providers

Any model, one surface

OpenRouter (default openai/gpt-4o-mini) routes hundreds of models — or go direct: Anthropic (Claude, prompt caching), OpenAI, Gemini, local Ollama, or your own.

BYOK

Your key, your spend

No proxy, no relay. Token budgets, usage tracking, and typed errors so cost stays visible and failures stay debuggable.

MCP

Plug in external tools

Declare any MCP server in config — GitHub, Postgres, your internal tools — and its tools become first-class agent tools in the autonomous loop. Zero-dependency client built in.

8×

Parallel subagents

Big tasks fan out to concurrent, context-isolated workers — only their summaries return, so a wide sweep costs a paragraph, not a context window.

Memory

It remembers your repo

The agent saves durable project facts as it learns them — conventions, gotchas, decisions — deduplicated and secret-screened, back in context every session.

v0.4.0

LiteRuntime today

Format repair for XML tags, exponential backoff for transient failures, CloudRuntime on the roadmap. Apache 2.0, source open.