BLOCKWAY · CODING AGENT · NOW OPEN

Codeway

Name: Codeway
Author: Blockway

A coding agent for your terminal, and a chat on the web — running Agens Pilot with a million-token context, in 廣東話 as a first language. Built so every session can be replayed and nothing runs outside the rails you set.

Terminal + Web
Agens Pilot inside
1M context
廣東話 native
Text + Vision
Sandboxed
Replayable

$npm install -g codeway

Node 20+ · macOS & Linux · Apple Silicon and Intel

Open the web app Read the docs

See it work

You describe the change. It reads, edits, runs the tests, and reports.

A simulated session, on a loop. Every step below is a real tool Codeway has — read, search, edit, run — each executed inside the sandbox and written to the session log.

codeway — ~/shop-api

›

Agens Pilot · 1Msandbox: strict · approvals: ask12.4k / 1.0M tokens

What it is

One agent, in your shell and in the browser.

Terminal-first

A full-screen agent in your shell: it reads the repository, plans, runs commands, edits files and verifies the result. Vim keys, mouse, a live token budget — or headless with codeway exec for scripts and CI.

And a chat on the web

The same model in the browser at codeway.blockway.io. Drop a screenshot and ask about it, have it search and read the web, or ask for a picture and get one generated inline.

廣東話 native · 1M context

Agens Pilot answers in 廣東話 as a first language and holds a million tokens of your codebase in view — with vision, so a screenshot of a bug is a valid bug report.

One account, both places

Sign in once. A turn on the web and a turn in the terminal draw from the same monthly allowance — the free tier included.

Built differently

What others bolt on, Codeway is built around.

Eight design decisions that make it safer to trust, easier to verify, and able to do more than answer.

The log is the session

Every event a session produces is appended to one log; the state is a fold over it. Any run can be replayed exactly, and a recorded session is already a regression test — replay it after a change and see what differed.

Two safety knobs, kept apart

A real OS sandbox (macOS seatbelt, Linux namespaces) confines what commands can read, write and reach; a separate approval policy decides what stops to ask you. Loopback-only networking lets the agent run your dev server without reaching out.

Plugins that cannot overreach

Plugins run under Node's permission model and get only the capabilities they are granted. Installing one never quietly widens what a session can do.

A real browser, not screenshots

Codeway drives a real browser over the Chrome DevTools Protocol — navigate, click, read the page, take a screenshot — inside the same session and the same safety rails.

Sub-agents in their own worktrees

Hand part of a task to a child agent. Each runs in its own git worktree, so parallel work never collides, and the result comes back into the parent's conversation.

Standing goals for long work

codeway exec --goal keeps opening turns until the model says the goal is done or a budget of turns, tokens or minutes runs out — with context compaction so a long task does not drown.

Evals and cost, built in

codeway eval runs a task suite and reports what passed and what each task cost, so a change to a prompt or a model is measured rather than felt.

Your editor, your keys

Speaks the Agent Client Protocol so an editor can drive it, ships a full vim mode in the terminal, and serves a local web UI with codeway serve --http.

Side by side

A typical coding agent, and Codeway.

	Typical coding agent	Codeway
Session record	A chat history to scroll back through	✓An append-only event log — replay any session exactly; recordings double as tests
Safety	Usually one allow / deny prompt	✓Two independent knobs: a real OS sandbox and a separate approval policy
Extending it	Plugins often run with full access	✓Capability-scoped plugins under Node's permission model
Browser	Screenshots, or none	✓A real browser driven over CDP
Parallel work	One thread	✓Sub-agents, each in its own git worktree
Long tasks	Stops when the turn ends	✓Standing goals with turn / token / time budgets
Measuring a change	By feel	✓Built-in eval runner with cost per task
Other agents	Works alone	✓Takes delegated tasks from Claude Code or Codex over MCP, each in its own git worktree
Where it runs	Terminal or web, one or the other	✓Terminal, web and your editor (ACP) — one account
Language & context	English-first	✓廣東話 native · 1M context · text + vision

Works with Claude Code and Codex

Keep your agent. Let it hand the simple work to Codeway.

Already working in Claude Code or Codex? Keep it. Codeway plugs in as an MCP server, so your main agent can pass the well-scoped jobs — adding tests, renames, boilerplate, mechanical edits, docs — to a Codeway session running Agens Pilot. The expensive model's usage is kept for the work that needs judgement.

1Your agent decides
2Codeway does the work
3You review the diff

Claude Code · your main agent

›

● Well-scoped, and a test run can prove it → delegate

codeway_delegate(task, check="pytest -q")

tests/test_slugify.pyreview

+ def test_spaces_become_hyphens():
+ def test_lowercases_input():
+ def test_empty_string():
  … 4 more · pytest -q → 7 passed

Apply

MCP

Codeway · Agens Pilot · own git worktree

✓ worktree started from your working tree

Main model

Agens (Codeway)

Your main model judges and reviews. The legwork runs on Agens, in your Blockway allowance.

1
Connect it
Run one setup command for Claude Code or Codex. It registers Codeway's MCP server, lists every change and asks before writing anything.
2
Your agent delegates
When a job is simple and clearly bounded, your agent hands it to Codeway as a complete instruction — optionally with a check command that proves it done. For Claude Code, a bundled skill says when to delegate and when to keep the work.
3
You review the diff
Back come a summary, the changed files, the diff, the proof and the Agens tokens used. Apply it to your working tree, or discard it.

Isolated by design

Each task runs in its own git worktree, started from your working tree as it is — uncommitted edits included.
Confined to auto-edit in a workspace-write sandbox, whatever your settings say. Anything that would need approval is refused.
Nothing in your checkout changes until you apply. If the files changed meanwhile, apply refuses rather than overwrite them.
Every task is an ordinary Codeway session: recorded, replayable, with its token use reported — never estimated.

A real run

Claude Code → Codeway

“Add pytest tests for slugify”

start to finish: 16 s
steps: 10
Agens input tokens: ~88k
Agens output tokens: 1.4k

✓7 tests written and applied — all passing.

For Claude Code

$codeway setup claude-code

For Codex

$codeway setup codex

Needs Codeway installed and signed in on the same machine.

Get started

Two minutes to your first turn.

1
Install
$ npm install -g codeway
Node 20+ · macOS & Linux · Apple Silicon and Intel. A prebuilt binary per platform — no compiler.
2
Sign in
$ codeway login
Opens your browser. The same Blockway account as the web app.
3
Start
$ codeway
Run it in a project directory and say what you want done.

Prefer the browser? Open the web app — nothing to install. The full manual is at codeway.blockway.io/docs.

Two surfaces, one account

Start in the terminal, continue on the web. Or the other way round.

›_

Terminal

codeway · TUI, exec, ACP

One Blockway account

one monthly allowance, shared

⌕

Web app

codeway.blockway.io

The model inside

Powered by Agens Pilot

Blockway's Cantonese-first open model — a 1,000,000-token context, text and vision, reliable tool calls — trained and evaluated against Codeway's real workloads. Open weights under Apache-2.0.

Explore Agens Pilot

Open today. Free to start.

Install the terminal, or open the web app — the same account and the same allowance either way. For a private deployment on your own hardware, talk to us.

Open the web app Talk to us