terminal-first coding agent

Your model knows how to code.Metis helps it finish.

Better context, reusable experience, and verified results around the model you already use.

npm i -g @wholiver_hu/metis@rc

01 / WHY METIS

Same underlying model.

A better way to
search, remember, work, and check.

Metis does not replace the model or change its weights. It strengthens the system around it—so the model begins with better evidence, preserves useful experience, and finishes with proof.

LIVE scroll-driven Metis session

metiscontext loaded

[Context]

AGENTS.md

docs/security.md

[Memory]

brain-map.md → auth-refresh-boundary

[Skills]

repository-search · verification

Build the auth flow. Reuse existing patterns and verify every requirement.

Checking project instructions and prior lessons…

CONTEXT BEFORE CODE

Give the model the right starting point.

Metis loads project instructions, relevant skills, and prior technical knowledge before substantive work begins. Better context changes the quality of every decision that follows.

AGENTS.md · skills · brain map

SEARCH BEFORE ACTION

Understand the repository. Then touch it.

Metis traces existing code, tests, ownership, and constraints before editing. That means fewer unsupported assumptions, smaller changes, and stronger compatibility.

repository search · authoritative sources

MEMORY & LESSONS

Experience compounds across sessions.

Useful decisions and hard-won technical lessons remain available after the session ends. The next task starts with evidence instead of rediscovery.

reusable knowledge · weighted recall

DREAM

Completed work becomes durable knowledge.

Dream reviews the full work history, filters routine noise, and promotes only genuinely reusable insights into structured memories and lessons.

consolidate · connect · clean up

VERIFIED COMPLETION

“Done” is an evidence-backed state.

Metis builds, tests, inspects output, and compares the result against every original requirement before reporting completion.

build · test · inspect · compare

ONE AGENT LAYER

Terminal-first. Not terminal-only.

Work in the interactive TUI, compose Metis in scripts with Print or JSON, integrate over RPC, or embed it with the Node.js SDK.

TUI · Print/JSON · RPC · SDK

02 / PRACTICAL OUTCOME

57^%

Less time in one same-task user comparison.

Metis1m 30s

Baseline3m 30s

One user test, not a universal benchmark. Results vary by task, model, tools, and environment.

03 / DESIGN PRINCIPLES

Reliable by habit,
not by hope.

Metis makes disciplined agent behavior part of the harness.

⌕

No action before search

Investigate code and constraints before making changes.

◇

No context left to chance

Load instructions, skills, and relevant experience up front.

↻

No lesson learned twice

Carry durable technical knowledge into later sessions.

≋

No long task without state

Append-only logs preserve decisions, errors, and next steps.

✓

No “done” without proof

Build, test, inspect, and check every requirement.

⌁

No single interface lock-in

TUI, Print/JSON, RPC, and SDK use the same agent layer.

open source · MIT licensed

Give your model
a better way to work.

View on GitHub ↗