Explorer
Codebase Recon
Terminal-native agent orchestration
Atelier
Run one prompt in a local repo and let the harness route the work through the right agents. You keep the terminal, the history, the approvals, and the final call.
Built for developers who already live in a terminal and want agents to work inside the repo boundary, not around it.
- 5
- runtime paths
- 8
- agent profiles
- 1
- active run
Agent Roster
One run, the right specialists.
The Orchestrator owns the run and delegates to focused agents — recon, knowledge, design, review, advice, planning, and implementation — each with its own capabilities and runtime.
Librarian
Knowledge Retrieval
Designer
Visual Excellence
Reviewer
Critical Review
The Orchestrator
Supreme Executor & Delegator
Oracle
Strategic Advisor
Consul
Planning & Reflection
Fixer
Implementation Specialist
Why it exists
Agent work needs a control plane.
Atelier treats agent activity as a run with routing, scopes, results, and evidence. The runtime can think and stream. The harness owns file edits, shell commands, approvals, limits, and the durable record.
Feature Map
Everything Atelier is already shaped to do.
The site is pulling from the current README and project docs. That means the boring but important pieces are here too: config, distribution, history, limits, and tests.
One prompt stays one run.
Run Orchestration
- Orchestrator-owned run plans
- Explorer, Oracle, Fixer, Reviewer, Consul, Librarian, and Designer profiles
- Clarifying questions when scope is not safe enough
- Serial council for high-risk architecture, security, and review calls
- Parallel step groups for disjoint file scopes
- Run, step, command, and review-fix limits
A working TUI, not a detached log.
Terminal Interface
- Chat for routing decisions, actions, diffs, diagnostics, and final results
- Agent Roster with model, runtime, availability, and active step
- Input Composer for prompts, answers, and interrupts
- Live runtime stream snippets while an agent is working
- Follow mode that respects manual scrolling
- Slash commands for goals, config, subtasks, agents, and skills
Use the CLIs and credentials you already have.
Runtime Choices
- Codex CLI runtime with Codex-owned login
- Claude CLI runtime with Claude-owned login
- Cursor Agent CLI runtime with Cursor-owned login
- Z.ai HTTP runtime through an environment credential reference
- Fake runtime for deterministic tests
- Per-agent model assignments and fallback chains
The harness owns the actions.
Control Plane
- Capability enforcement for read, edit, command, verify, and review work
- Action approval for high-impact operations
- Tool allowlists for narrower agent profiles
- Credential references instead of secrets in config
- Redacted TOML output from print-config
- VCS actions only when explicitly requested
Every run leaves useful evidence behind.
Project Memory
- Session events under .atelier/sessions
- Artifacts for large output, diffs, and run material
- Run records under .atelier/runs
- Context resume from harness-owned history
- Clean-session command for local history cleanup
- Codemap state for repository orientation
Install once, run where the work is.
Distribution
- Global npm package at @matheusbbarni/atelier
- Prebuilt native platform binary packages
- macOS, glibc Linux, and Windows arm64/x64 targets
- GitHub Release archives and checksum manifests
- Cargo install path for source builds
- atelier --update through the npm launcher
How a run moves
The loop is explicit.
No hidden handoff between agents. Each result comes back to the Orchestrator, and the next step is a decision the harness can show, persist, and bound.
- 01
Prompt
You type the work once. Atelier starts one run and records it.
- 02
Route
The Orchestrator chooses a single agent, a council, or a parallel group.
- 03
Act
Agents ask the harness for reads, edits, commands, approvals, and verification.
- 04
Review
Reviewer findings go back to the Orchestrator instead of editing directly.
- 05
Record
History, artifacts, diffs, diagnostics, and run results stay under .atelier.
Runtimes
Bring the agents you actually use.
Codex, Claude, and Cursor run through their local CLIs and login state. Z.ai uses an API-key environment reference. The fake runtime keeps tests deterministic. Atelier keeps those paths separate so config stays readable and credentials stay out of project files.
Product Views
Four views of the same control loop.
The UI is not trying to hide the work. These real TUI states show the run surface, pickers, and command reference inside the product.
Run Surface
Agents, chat, and the prompt composer stay visible in one terminal workspace.
Skill Picker
/skill: opens cached project and personal skills without leaving the run.
Agent Picker
/agent: shows enabled agent profiles, runtimes, and capability sets before routing work.
Help Overlay
/help keeps TUI and CLI commands available while the session stays underneath.
Command Surface
The flags are part of the product.
The CLI is intentionally small. You can open the TUI, inspect setup, print config, manage codemaps, clean history, and update the npm install without learning a second tool.
CLI
atelier- Open the TUI in the current working directory.
atelier --cwd <path>- Start from a selected working directory.
atelier --config <path>- Use a specific harness configuration file.
atelier --doctor- Check runtime setup before work starts.
atelier --doctor --json- Emit doctor output for scripts and CI.
atelier --print-config- Print the merged config with secrets redacted.
atelier --init-config- Create starter config and instruction files.
atelier --codemap init- Write folder-level codemap docs and state.
atelier --codemap changes- Report stale codemaps without writing.
atelier --codemap update- Regenerate stale codemaps and refresh state.
atelier --clean-sessions- Delete project-local session history after confirmation.
atelier --clean-sessions --yes- Clean project-local history without a prompt.
atelier --update- Update the global npm install.
atelier --debug- Launch with extra diagnostics.
TUI
/goal <text>- Set the active session goal.
/goal- Show the current goal.
/goal clear- Clear the active goal.
/config- Show active config files, preset, and warnings.
/subtask <agent> <task>- Run a bounded child task through one enabled agent.
/agent:<name>- Route a prompt to a selected agent profile.
/skill:<name>- Prefix a prompt with a cached project or personal skill.
/reload:skills- Refresh cached skill names.
Install
Start with npm. Build from source when you need to.
The npm package installs a prebuilt native binary and leaves your home config alone. Run doctor first, then open Atelier in the repo you want agents to work on.
npm install -g @matheusbbarni/atelier
atelier --doctor
atelier
Source build: cargo install --path .