Cezar — AI Coding Agent Orchestrator
One control center to run Claude Code, Codex, OpenCode and other AI coding agents in parallel — locally or 24/7 on your own server.
- Source repo
- open-mercato/cezar
- Stars
- ★ 1.1k
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 62/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free, no paid service needed
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A solo developer who wants several coding agents working on different tasks across repos simultaneously, each in its own worktree.
- Not a fit if
- Teams that want a chat-first agent inside Slack or IM rather than a web cockpit
- Projects unwilling to let agents modify the repo via git worktrees
- Website
- cezar.run ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release desktop-v0.1.12 (2026-11-20)+10.6 ★/day over 7 days
- Source review
- 62/100 · Some gaps
What does this agent do, and when should you use it?
Cezar is the open-source orchestrator and ADE from the open-mercato/cezar repository, providing a single web cockpit for running multiple AI coding agents concurrently. It reuses your existing logged-in claude, codex, opencode or pi CLIs, so no API keys are required. Every task runs in its own git worktree, letting several agents work simultaneously while extra tasks wait in a queue. Workflows are short YAML files and skills are Markdown files, with agents mixable per step; live runs stream agent text, tool calls, tokens and cost. It has no database — all state is plain files under .ai/cezar/ — and can be started locally with npx cezar-run or deployed to an Ubuntu VPS via server-install for 24/7 operation, reviewable from a fully responsive phone UI.
Cezar takes task descriptions (typed, with attached files, or launched from a GitHub issue / Jira / Linear) and runs them in a fresh git worktree according to a YAML workflow (such as the built-in quick-task), executing agent steps plus shell checks. A typical workflow:
yaml
name: fix-and-verifysteps:
- id: implement
prompt: "{{task}}"
skill: project-conventions
runner: codex
- id: verify
command: "npm test"
onFail: { retry: implement, max: 2 }It invokes your locally logged-in agent CLIs to execute steps, retrying automatically with the error fed back when a check fails. The cockpit streams every step, tool call, token and cost live. You can run the same task ×2 or ×3 and compare diffs, agents can delegate child tasks (up to four children in flight per parent, sharing the parent budget), and Automations launch tasks on schedules or from GitHub/tracker events. Everything — workflows, skills, task state — is stored as plain files in .ai/cezar/. Results are reviewed as diffs, with notes or draft PRs; nothing auto-merges.
- A solo developer who wants several coding agents working on different tasks across repos simultaneously, each in its own worktree.
- A remote/freelance developer deploying Cezar on a VPS so agents keep working 24/7 after the laptop closes, reviewing diffs from a phone.
- An open-source maintainer handing a GitHub issue to an agent in one click and reviewing the resulting draft PR.
- Teams on Jira or Linear launching workflows directly from tracker issues and configuring event-driven automations.
- A developer comparing approaches by running the same task ×2 or ×3 and keeping the best diff.
- A team tracking per-project token usage and reported cost through the Usage & cost dashboard.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npx cezar-runnpx cezar-run server-install --platform ubuntu-vpsnpx cezar-run run "add a -- flag to the export command"npx cezar-run init
- Ports
4321- Runtime
- Node.js 20+
- Install notes cover
- macOS · Linux
How do you install or deploy this agent?
Prerequisites: Node 20+ and at least one logged-in agent CLI (Claude Code, Codex, OpenCode, or pi); git and gh are optional. No database or API keys needed.
bash
cd your-repo
npx cezar-runThis opens the cockpit at http://localhost:4321. To try it without logging in:
bash
CEZ_DRY_RUN=1 npx cezar-run(Uses a built-in mock agent.) To set up a server:
bash
npx cezar-run server-install --platform ubuntu-vpsThis sets up HTTPS, a login and a system service so the cockpit is reachable from anywhere, including your phone. A macOS + ngrok guide is also documented.
How do you use this agent?
Type a task in the cockpit, pick a workflow and hit Start, or run headless from the CLI:
bash
npx cezar-run run "add a -- flag to the export command"Scaffold project configuration:
bash
npx cezar-run initTry the nightly build:
bash
npx cezar-run@nightlyCustom workflows are YAML files in .ai/cezar/workflows/; skills are Markdown files in .ai/skills. Workflows can be built by drag-and-drop in the UI and are saved as YAML. Each step can target a different agent via the runner field, and command steps pass when they exit 0. You can also connect GitHub, Jira or Linear to launch tasks from issues, or schedule recurring work and event triggers in Automations.
What are this agent's strengths and limitations?
- Reuses your existing agent CLI logins — no API keys — to drive Claude Code, Codex, OpenCode or Pi.
- Per-task git worktrees give genuine multi-agent parallelism, plus parent/child task delegation (up to four children in flight).
- Live streaming of agent text, tool calls, tokens and cost, with per-project cost comparison in Usage & cost.
- Automations combine schedules with GitHub/tracker event triggers, and VPS deployment enables 24/7 operation.
- No database: all state lives as plain files under .ai/cezar/, easy to inspect and back up.
- Requires Node 20+ and an already-logged-in agent CLI (claude, codex, opencode, or pi); without one, no real tasks can run.
- Parents are limited to four children in flight, and children share the parent's budget, which may constrain heavy parallel use.
- Agent usage bills against your own Claude/Codex/OpenCode subscriptions — the software is free but heavy use costs you tokens.
- Complex behavior depends on hand-written or drag-built YAML workflows and Markdown skills, which imposes a learning curve.
- A promoted cloud sandbox (openmercatocloud.com) exists, but its pricing is not stated in the source material reviewed.
How does this agent compare with similar options?
The README positions Cezar as an orchestration layer above the individual coding agent CLIs — Claude Code, Codex, OpenCode and Pi are the execution backends it invokes, while Cezar adds parallel dispatch, worktree isolation, workflows, automations and one cockpit, which none of those CLIs provide alone.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Cezar — AI Coding Agent Orchestrator This agent | 62 · Some gaps | CLIFree | ★ 1.1k | today | TypeScript | Codex · Claude Code |
| Vicoa | 67 · Some gaps | Desktop appFree + model costs | ★ 544 | 1d ago | Python | Codex · Claude Code |
| gnhf (Good Night, Have Fun) | 55 · Major gaps | CLIFree + model costs | ★ 4.3k | 4d ago | TypeScript | Codex · Claude Code |
| Kandev | 74 · Some gaps | CLIFree + model costs | ★ 1.2k | today | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
AGENTS.md and the README describe isolated worktrees, a human review gate, no automatic merge, explicit telemetry consent, secrets kept out of agent environments, and loopback binding by default. These support ordinary-use safeguards. Deductions reflect that the product can run arbitrary agent and shell workflows, Autonomous explicitly avoids stopping to ask, and VPS or external-proxy use widens access; the packet does not show the complete authorization model or data-flow implementation. AGENTS.md also records pending approval for default network behavior.
The package manifest, CI workflow, and package tests show maintenance mechanisms for typechecking, builds, unit suites, cross-platform webhook coverage, and release-package checks; the guidance also describes graceful degradation when dependencies are absent. Deductions reflect that these files do not prove CI passed, two star-count descriptions in AGENTS.md conflict on frequency, caching, and token behavior, and the packet lacks enough runtime error-handling code or a dependency lockfile to assess availability and failure messages fully.
The README explicitly covers multiple coding agents, terminal and browser use, mobile, local and VPS deployments, workflows, GitHub issues, and Jira/Linear; the CLI also offers headless and project-level entry points. Deductions reflect that OpenCode is marked experimental and the packet does not fully enumerate runner, platform, or permission boundaries; automation trigger details are summarized rather than fully specified.
Documentation entry points, quick start, CLI examples, and links to architecture and installation guides are clear; the MIT license text is complete, and the security reporting route, supported versions, and response timelines are stated. Deductions reflect a legacy CLI compatibility note, support limited to the latest minor line, no changelog content in the packet, and no verified publisher registry identity or explicit maintainer-team commitment despite the license attribution.
The product brings multiple coding agents, worktrees, queues, composable workflows, live status, diff review, and draft PRs into one cockpit, with reviewable output and CLI/API access. Deductions reflect that outcomes, correctness, and real cost control cannot be established statically; README cost and token figures are agent-reported, and autonomous execution or a persistent VPS adds supervision and resource costs.
The README's main workflow can be cross-referenced with the CLI entry point, API contract, CI, and package-test examples, while some security claims have corresponding descriptions in AGENTS.md or SECURITY.md. Deductions reflect that many feature and security claims remain assertions in README/AGENTS.md, test files show test design rather than execution results, and conflicting star-count descriptions in AGENTS.md reduce claim consistency.
- Autonomous mode is described as not asking during a run; before using it for long-running or VPS tasks, check the workflow, agent permissions, network access, and task scope.
- The README says VPS installation configures login and HTTPS, while CLI help warns that cezar has no built-in authentication and an external proxy must provide TLS and authentication; verify the actual bind address and proxy configuration before deployment.
- AGENTS.md contains conflicting descriptions of the star-count request's default state, frequency, cache, and token behavior; this static packet cannot establish the implementation.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
desktop-v0.1.12 cezar desktop v0.1.12
Release notes excerpt
Desktop shell v0.1.12. Installers below; the shell installs and updates cezar itself from the cockpit.View on GitHub ↗ -
Release notes excerpt
✨ Features - ✨ The cockpit is installable: add it to a phone or desktop home screen and it opens standalone, including behind a Basic auth gate. Live data stays live, there is no service worker. (#1762, #2038) *(@patzick)* - ✨ Managers: create global agents in the new Manager mode, connect repositories (read-only or write per repository), and chat with them. A manager refreshes on a cadence (a check with nothing new makes no model call), runs saved jobs, starts and steers tasks, reports every task waiting on you, asks decisions on question cards, and proposes merges, automation enables and …View on GitHub ↗ -
Release notes excerpt
Highlights Four new agent runners (Gemini CLI, Kimi Code, Antigravity and omp) and HTML artifacts, with Assess Impact for PRs and specs as the first feature built on them. Onboarding installs an agent and signs you in without leaving cezar, a task that runs out of usage can fall back to another provider or account, and the task header shows live Claude usage. Jira and Linear tokens are now encrypted and tracker automation agents no longer receive them, so read Breaking before upgrading. The sidebar groups projects into folders and folds into an icon rail. ⚠️ Breaking …View on GitHub ↗
FAQ
Do I need to pay or provide API keys?
Will agents merge code automatically?
What happens when a task fails?
Can it run fully offline?
Where is my data stored?
Which port does Cezar — AI Coding Agent Orchestrator use?
Is Cezar — AI Coding Agent Orchestrator open source?
Is Cezar — AI Coding Agent Orchestrator safe to use?
Is Cezar — AI Coding Agent Orchestrator still maintained?
Who makes Cezar — AI Coding Agent Orchestrator?
What is Cezar — AI Coding Agent Orchestrator's official website?
Thanks — we use this to improve the page.