Get your agents under controlResearch · Innovate · Plan · Execute · Review · Test & ship

Give your AI agent a bench to work on.

Research, Innovate, Plan, Execute, Review, then test and ship. Your agent writes the code; Workbench makes sure the work gets done properly, on your machine, with you in control.

Download Talk to us

Workbenchv1.1.1-beta.0

workbench/workbenchmainEXECUTEreading agent-providers.mjs

claude-opus ▾Stop

Files · 5

▾server
agent-providers.mjs+42 −6
config-renderer.mjs+18 −2
schema.sql+4 −0
▾client/src/workbench/pages
SessionPage.tsx+64 −21
▾client/src/workbench/lib
diff.ts+9 −0

DiffFull fileserver/agent-providers.mjs

@@ -112,6 +112,8 @@+3 −1

112 export function buildAgentSpawn(session, provider) {

113 const env = augmentedPath(process.env);

114− const model = provider.defaultModel;

114+ const model = session.model ?? provider.defaultModel;

115+ // model changes apply on the next turn, not mid-run

115 const child = spawn(provider.bin, args, { env });

116 const child = spawn(provider.bin, args, { env });

117+ runningProcess.set(session.id, child);

116 return { child, model };

118 return { child, model };

117 }

119 }

@@ -112,6 +112,8 @@+3 −1

112112 export function buildAgentSpawn(session, provider) {

113113 const env = augmentedPath(process.env);

114− const model = provider.defaultModel;

114+ const model = session.model ?? provider.defaultModel;

115+ // model changes apply on the next turn, not mid-run

115116 const child = spawn(provider.bin, args, { env });

117+ runningProcess.set(session.id, child);

116118 return { child, model };

117119 }

Reasoning

Why this changed

The session model was read from the provider default, so switching models mid-run had no effect on the next turn.

agent-providers.mjs

Fix

Spawn now prefers session.model and tracks the child process so Stop can reach it.

L114–117

Reply with this file's diff…

Context: server/agent-providers.mjs · diff · 2 reasons

00 · Overview

01

A process the agent can't skip

Every session moves through Research, Innovate, Plan, Execute, and Review. The agent looks before it leaps, and nothing is committed until you approve the plan.

02

One rulebook. Every agent.

Your rules, your project memory, and the session focus go with every agent you use. Plan with one and build with another without losing the thread.

03

Code review that shows its work

Built for agent-sized pull requests: walk through the change, compare fixes before you pick one, and check that feedback was really fixed. Not just a wall of +143000/−67000.

01 · Start// pipelines

From an empty folder to a real start.

Starting from scratch? Two starters set up the plan for you. Describe what you want once, and the session follows it.

01
Website builder
Describe the site and get a complete marketing build: pages, structure, and something ready to put online.
02
Project docs
Kick off a full project plan: discovery, sprints, user stories, and the docs your team needs before feature work.

02 · Session// sessions

From first question to merged PR, with you in charge.

Give Workbench a repo and a goal. The agent researches, weighs options, writes a plan for you to approve, and only then builds. It stays on topic, remembers your project, and stops to ask when a decision is yours.

See everything a session can do

Bench mode

01RESEARCH

The agent maps your code and the problem before it touches anything.

Problem understood

02INNOVATE

It weighs a few ways to solve it and talks them through with you.

Approach chosen

03PLAN

A written plan and checklist you approve. Commit stays locked until you do.

Plan approved

04EXECUTE

One click and the agent builds it on its own branch, then drafts the commit message.

Code written

05REVIEW

The agent checks its own work, and a quality check flags leftover secrets, debug code, and gaps.

Work checked

06TEST & SHIP

Work through a short test checklist, then commit, push, and open the pull request from the same screen.

PR open

Loop

Hands-off. Set a goal, turn Loop on, and send. The agent keeps going, stops only for your questions, then hands you a checklist to test.

Learn more

acme/webIdleIdle

Loop offMesh off

RESEARCH

Mapped the login flow and the token store.

PLAN

Plan written and accepted automatically.

EXECUTE

Building on feature/2026_10_08_token-rotation.

REVIEW

QA passed. Ship gate open.

Go test

Loop complete. Tokens now rotate on every login.

Log in twice and check the old token is rejected
Refresh the page and stay signed in
Check the audit log shows the rotation

Tests fail or need more work? Send a message below. Loop resumes.

Looping

Queue a follow-up · it sends when this turn finishes

Break loop

Type your message…

Mesh

Many angles at once. Run research, a code review, or a security check in parallel and get one clear set of findings before any code changes.

Learn more

·

Ready to start

Pick a review type and press play.

0/5 steps

Code review ▾2 at once

1
Correctness
Tests, edge cases, regressions.
Waiting
2
Security
Auth, input, secrets, attack surface.
Waiting
3
Performance
Speed, memory, and scale.
Waiting
4
Summary
One ship or no-ship recommendation.
Waiting
5
Final check
Filters weak or duplicate findings.
Waiting

Findings

Findings appear here after the final check step.

Implement findingsContinue in chat →

Loop on. Building from the findings.

03 · Context// context

Know what the agent is carrying before the run gets expensive.

Long sessions fill up fast. Workbench shows how full the agent's memory is, straight from the agent's own numbers, plus what the session has used and roughly what it has cost, so you know when to trim the scope or start fresh.

// without it

You only find out the window was full after the answers start drifting

Usage is buried in a terminal, if your agent shows it at all

A multi-step run feels like one message until the bill arrives

// with Workbench

See how full the session is after every step, from the agent's own numbers

Track total usage and an estimated cost across the whole session

Know when to narrow the scope or start a fresh session

04 · Knowledge// knowledge

Memory you can see.

Explore Knowledge & Brain

Your project notes, a map of how everything connects, and an agent that gets the right few notes for each step, not everything you ever wrote. Finished sessions add what they learned automatically.

Notes

Read, correct, and keep your project's memory. It's yours, on your machine.

Map

Brain shows how your services, standards, and code connect.

Right notes, right time

Each step gets the few notes that matter, not the whole collection.

05 · Review// review

PR review that shows its work.

Agents can write thousands of lines in minutes. Workbench helps you understand them: what changed, why, and what could break. Then it helps you act: post comments, compare fixes, send issues back to the agent, and check that each one was really closed.

See the full review toolkit

See the change, not just the files

Every pull request opens with the description, the changes, and the issues found, ranked by how much they matter. Nothing is posted to GitHub until you say so.

Ask why it changed

A guided tour of the pull request, step by step. Ask the agent about any step, check the work against your own acceptance criteria, and turn an answer into a comment.

A stack of PRs, reviewed as one

Review a series of related pull requests as a single change. Each comment lands on the PR that owns the line, and you can approve and merge the whole stack in one go.

06 · Ship & Ops// ops

Run it, test it, ship it, from the same session.

Start the app for the branch you're on, test it for real, then ship from the same session. No separate ops maze.

Run the stack for this PR

Start, rebuild, or stop the services for this session's branch, so you can actually run what the pull request changes.

Services

07 · Agents// agents

One rulebook. Every agent CLI.

Cursor, Claude Code, Codex, and OpenCode all run inside Workbench with the same rules and skills. Use a different agent for planning and building if you like; the plan and progress follow along. Workbench runs the process. It is not another code editor.

Become fluent withCursor Agent

Cursor Agent

Rules, skills, and tools, delivered automatically

Claude Code

Rules and skills, delivered automatically

Codex CLI

Rules and skills, delivered automatically

OpenCode

Rules and skills, delivered automatically

// also on the bench

The rest, without the clutter.

Plan approval before commit

Commit stays locked until you approve the plan and the quality check is clean.

History and recovery

If the app closes mid-task, pick up where you left off from History, or clear out old work.

Steer it from Slack

Start sessions, answer the agent, and act on reviews from Slack on your phone. Desktop notifications cover you when you're at your desk.

Rules and skills that stick

Shared defaults plus your own rule packs, loaded only when needed so sessions stay fast.

Profiles and accounts

Separate workspaces per person, each with the GitHub accounts they link.

Questions as clear choices

When the agent needs a decision, you get a card with options to click, not a question buried in a wall of text.

// for developers

Built to help you ship, not to watch how you work.

We do not collect metrics or usage analytics. Workbench exists to help you build: no account, no cloud sync, no tracking. Create a profile on first run and start working. Agent calls go only to the provider you chose.

Accounts required: None

Data storage: On your machine

Telemetry: None. We do not collect metrics

GitHub linking: Optional

Talk to us

Want help getting set up?

Downloads are live on the Download page. Book a short call if you want help getting set up while the first builds settle in.

You get

Builds for Linux, Windows, and macOS
A direct line for bugs, crashes, and feature requests
Influence over the roadmap while it is still soft

We ask

You already use an AI coding agent day to day
You are comfortable with git and pull requests
You will tell us when something is broken or confusing

Talk to us

On the roadmap

More agents: Support for more coding agents and smoother runs across machines.
Team workflows: Share rule packs and move profiles between machines more easily.
Richer review: Review templates, batch PR review, and tighter checks before you merge.
Cross-platform polish: Smoother installs and updates on every platform.
Mesh library & budgets: Browse saved Meshes from Repos, push findings into Review, and set turn or token budgets beyond the parallel worker cap.
Semantic knowledge: Smarter search across your company and project notes, so the agent finds the right note even when the wording is different.

FAQ

Questions, answered.

Downloadv1.2.0-beta.1

Ship with your agent, not around it.

Get your agents under control. Two minutes from install to your first structured session, all of it on your own machine.

Download Talk to us What's in 1.2.0-beta.1

Give your AI agent a bench to work on.

A process the agent can't skip

One rulebook. Every agent.

Code review that shows its work

From an empty folder to a real start.

Website builder

Project docs

From first question to merged PR, with you in charge.

Loop

Mesh

Know what the agent is carrying before the run gets expensive.

Memory you can see.

Notes

Map

Right notes, right time

PR review that shows its work.

See the change, not just the files

Ask why it changed

A stack of PRs, reviewed as one

Run it, test it, ship it, from the same session.

Run the stack for this PR

Docker Compose

Ops log

One rulebook. Every agent CLI.

The rest, without the clutter.

Plan approval before commit

History and recovery

Steer it from Slack

Rules and skills that stick

Profiles and accounts

Questions as clear choices

Built to help you ship, not to watch how you work.

Want help getting set up?

Questions, answered.

Ship with your agent, not around it.