Atris Project Workbench
Give coding agents shared project memory, bounded work, and runnable checks that show what they completed.
- Source repo
- atrislabs/atris
- Stars
- ★ 71
- Last updated
- 1d ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 47/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code
- Cost
- Free tier plus a paid hosted plan
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Development teams that want Claude Code, Codex, or Cursor to share repository memory across sessions.
- Not a fit if
- Teams that require all task state to stay local
- Teams that need member changes to land automatically without GitHub auto-merge
- Website
- atris.ai ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v3.90.0 (2026-11-20)
- Source review
- 47/100 · Major gaps
What does this agent do, and when should you use it?
Atris is a Node.js CLI workbench for code repositories, with an MCP server and optional cloud workspaces. After initialization, it keeps shared context in repository files, including agent instructions, a navigation map, tasks, logs, and wiki pages. Agents can follow a plan, do, review loop or advance work through run and autopilot. Tasks can carry verification commands, and work can produce receipts; member work also gets a review from a separate engine. Cloud workspaces, business features, and integrations require additional setup, and the README also documents ax, a chat CLI that uses hosted Atris cloud by default.
Running atris init in a project creates an atris/ workspace; atris then loads context and hands the workflow to the instructions in atris/atris.md. Teams can create work with atris task new, attach a --verify command, and manage task state with next, ready --proof, and human accept. atris wiki, atris ingest, and atris loop organize source material, maintain linked pages, and refresh wiki status. Use atris run or atris autopilot to advance work, or run member jobs in separate copies with verification and review. Atris records activity in logs, task events, and receipts. In cloud business workspaces, Supabase holds task state and Swarlo provides the live claim layer.
- Development teams that want Claude Code, Codex, or Cursor to share repository memory across sessions.
- Maintainers who need to require a test or other deterministic check for each agent task.
- Teams that want coding agents to work in parallel on safe tasks, with independent review and branch-based delivery.
- Engineering leads who want to turn loose local source material into a linked project wiki with status.
- Teams that need task tracking, run history, and operational recaps in a business workspace.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
npm install -g atris
- Runtime
- Node.js 18
How do you install or deploy this agent?
Requires Node.js 18 or later; atris task requires Node.js 22 or later. Install the CLI globally:
npm install -g atris
atris --versionRun atris setup for cloud workspaces, business features, or integrations. The README does not provide MCP client configuration steps or required credentials.
How do you use this agent?
Initialize the workspace in the target repository, then start Atris:
cd your-project
atris init
atrisAgents should read atris/atris.md and follow its instructions. Example task with a verification command:
atris task new "fix the flaky login test" --verify "node --test test/login.test.js"After work, submit evidence with atris task ready <id> --proof "..."; a person can then run atris task accept <id>. To start one bounded objective, run:
atris run "fix the flaky login test"What are this agent's strengths and limitations?
- Project memory lives in ordinary repository files that agents and people can share and review in Git.
- Tasks can require runnable verification, and member work receives a review from an engine other than its builder.
- Installed coding CLI engines can be switched and assigned by job role.
- The workflow supports local wiki, logs, and tasks, with optional cloud sync and business workspaces.
atris taskrequires Node.js 22; the base installation requires Node.js 18 or later.- The README does not explain MCP client setup, so the connection steps need confirmation before adoption.
- Automatic member landing requires self-landing to be enabled, with GitHub auto-merge and CI handling the merge.
- Cloud business workspaces use Supabase for task state and Swarlo as the live claim layer.
How does this agent compare with similar options?
The README contrasts Atris project memory with Mem0, Zep, and Letta, which it describes as remembering chat-app users; Atris stores shared project memory in repository files.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Atris Project Workbench This agent | 47 · Major gaps | CLIFreemium | ★ 71 | 1d ago | JavaScript | Codex · Claude Code |
| no_human | 75 · Some gaps | CLIFree + model costs | ★ 322 | 1mo ago | Python | Claude Code |
| Citadel | 84 · Good | Agent plugin / skillFree + model costs | ★ 935 | 16d ago | JavaScript | Codex · Claude Code |
| mex Project Memory | 84 · Good | CLIFree | ★ 1.8k | 2d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence shows real trust controls: the MCP campaign_ask/campaign_answer tools explicitly stop before irreversible actions (sending, paying, deleting, deploying, publishing) and an agent cannot close an ask; the tune skill requires a diff and explicit approval; autoland is off by default and protected lanes (payments, deploys, credentials) never auto-land. Deductions are equally concrete: CLAUDE.md forces `atris atris.md` execution and output display on every conversation start with no confirmation; the swarlo skill curls localhost:8090 with a Bearer KEY and instructs `git pull && git push` automatically; README's `atris agent-client create` mints client secrets with scopes like mcp:read and x-search without in-package explanation of how scopes map to least privilege; data flow is not centrally documented (ax defaults to hosted cloud, design tools call an external billed API); sensitive data handling is one line about `atris wiki --private` under .atris/presidio/ with no encryption, retention, or log-redaction detail; rollback has `member tidy`, `land --reap`, `run prune-logs` but no unified rollback for landed or pushed changes; source attribution has MIT and an author field but the publisher is unverified and many external links (atris.ai blog, hosted API) cannot be checked in-package. Dependency security is a strength (zero runtime deps, npm ci, read-only CI permissions) but the prepare script writes git hooks at install time, so not full marks.
Reliability evidence: CI shards node --test across 8 jobs plus a macOS sandbox workflow; accept-verify-gate tests pin exact failure text (missing directory vs command did not pass vs non-falsifying command), showing failure messages are deliberately designed. Deductions: Node version requirements contradict across README (18+), README task note (22+), and package.json engines (>=22.13); the README is very long and asserts runtime behaviors (fleet scoring, landing rules) that cannot be confirmed against code statically; dependency availability is strong on the zero-dependency side but ax and design tools depend on a hosted cloud service with no documented offline degradation path.
Audience and scenario coverage is broad: README addresses individuals, businesses, AgentXP players, and business ops, and AGENTS.md/CLAUDE.md/FOR_AGENTS.md target different agent tools. Deductions: capability boundaries are blurry, since the README simultaneously claims to be a CLI, an MCP server, an operating system, a game, a payments system, a social Host, and a site-publishing platform without a clear scope statement; trigger precision is weak because several skills (tune, swarlo, endgame, aeo) overlap and some depend on local paths (~/arena/empire/swarlo, /workspace/atris/...) tied to a specific machine layout; environment fit is partial because macOS sandbox tests run separately and Linux CI skips them, implying some security boundaries only hold on one platform with no documented behavior elsewhere.
Information architecture is clear: README has a Critical Files table, a directory tree, and a Core Commands table; AGENTS.md has Core Files and Proof Surfaces tables. Install notes are complete (npm install -g, Node version, atris init). Examples are plentiful. License is full MIT text with a named copyright holder, earning full marks. Deductions: naming stability is weak, with README itself noting dispatch/go are aliases of run and up of wake, plus `--legacy` retaining old chains; known limitations are scattered (Node version, autoland opt-in) with no consolidated limitations section; version numbers disagree between package.json (3.82.14) and marketplace.json (1.0.0) and there is no CHANGELOG; maintenance responsibility is only an author field and issues link, with an unverified publisher and no security disclosure channel or SLA.
Output usability is decent: atris recap, receipt, run logs, and task --json produce pasteable and machine-readable artifacts, and MCP tools return structured JSON with credits_charged and balance. Deductions: marginal value is questionable because the advertised 'memory' is plain text files in the repo (MAP.md, wiki/, logs/) that any agent can read and write directly, so Atris's increment is mainly workflow orchestration and verification gates, which are described but not benchmarked in-package; cost-benefit is unclear because design tools bill per call (10/20/1 credits), campaigns default to a $20 budget, and aeo drafts cost 5-20 credits, yet there is no documented spend cap, budget-exhaustion behavior, or refund on failure.
Verifiability is weak: README and AGENTS.md contain many quantitative claims ('803 receipts, 1711 episodes', '670/674 accepted', '24 of 28 bounces') with no reproducible data file or query command in the packet; the fleet scoring rules are described in extreme detail but cannot be checked against code statically; external links (atris.ai blog, aeo check page) cannot be verified in-package. Positives: accept-verify-gate tests pin failure text and decision logic, which is traceable, and AGENTS.md clearly separates 'agent proof ready' from 'human accept' as two gates, giving some fact/inference separation. Overall the bridge between claims and evidence is thin, so each criterion scores 1.
- CLAUDE.md forces `atris atris.md` execution and output display on every conversation start with no user confirmation; confirm whether this can be disabled before adoption.
- The swarlo skill hits localhost:8090 with a Bearer KEY and instructs automatic `git pull && git push`, which can cause unreviewed pushes in multi-agent setups.
- Node version requirements contradict across README (18+), README task note (22+), and package.json engines (>=22.13); treat package.json as authoritative and verify before install.
- ax defaults to hosted cloud and design tools call an external billed API, with no central documentation of what data leaves the machine or how to disable or degrade.
- Design tools bill per call and campaigns default to a $20 budget, but there is no documented spend cap, budget-exhaustion behavior, or refund on failure.
- Publisher identity is unverified, and many external links and quantitative claims in the README cannot be checked in-package; treat them as unverified assertions.
- The package.json prepare script writes git hooks at install time, an install-time side effect that should be evaluated in a controlled environment.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v3.90.0 Atris 3.90.0: your repo remembers its mistakes
Release notes excerpt
Your project keeps a short list of the mistakes it has made before. Every change gets checked against each one. - atris code-review reads atris/SPEC.md, gives each mistake type its own reviewer and refuses a change when one asks for changes. atris code-review init writes a starter list. One rule needs no model: a change cannot add more test lines than code. - Tell your coding agent "Set up https://github.com/atrislabs/atris". It follows FOR_AGENTS.md and ends with atris doctor --live, which asks a real model for one exact reply. …View on GitHub ↗ -
v3.83.2 Atris 3.83.2: the feed as a shared room
Release notes excerpt
The feed is a shared room for you and your team. - One command gives any project folder a team that posts ideas for you to approve: atris feed start <folder> - Team members post short notes, ask questions and answer each other. - You can post a note or a question yourself: atris feed say "<text>" [--ask] - React to any card and reply on it: atris feed react, atris feed reply - Approved writing work on an Atris cloud model comes back as a result. The same feed is in Atris Desktop 0.1.76. Install: npm install -g atrisView on GitHub ↗ -
v3.80.0 atris 3.80.0
Release notes excerpt
atris sentinel runs a full security audit of a repo. - It maps every source file and kind of attack worth checking, for free, before any model is called. - Hunters read the code and report what they checked. Every suspected hole goes to a different model that tries to prove it wrong, and a cited line that does not exist is thrown out first. - Runs add up. Areas whose code did not change stay covered, old findings get re-checked, and the next run spends its budget on what is left. …View on GitHub ↗
FAQ
Does Atris require one model provider?
Which Node.js version do I need?
atris task uses built-in SQLite and requires Node.js 22 or later.How does Atris decide a task is done?
--verify command. Member jobs run a check and receive review from another engine; human acceptance uses atris task accept.Do I need a cloud workspace?
atris setup.Is Atris Project Workbench free?
Is Atris Project Workbench open source?
Is Atris Project Workbench safe to use?
Is Atris Project Workbench still maintained?
Who makes Atris Project Workbench?
What is Atris Project Workbench's official website?
Thanks — we use this to improve the page.