atris
atris: you say what you want in plain words. atris builds it, checks it, and shows you proof.
Under the hood it is an operating system for AI work. It turns any repo into a workspace an agent can actually operate: shared context it reads before it acts, a plan -> do -> review loop, durable tasks and daily logs, and verification that ends work on a real check instead of a promise.
Every owner gets persistent computers to run it:
Owner = User | Business
Owner has many Computers
Computer = workspace + files + tools + secrets + memory + agents + validation loop
A computer can be personal or shared by a business. It can run a job like code, research, CRM, reporting, event ops, support, or business ops.
Memory for coding agents
Atris gives Claude Code, Codex, Cursor and any other agent a memory that lasts across sessions. The memory is plain files in your repo: atris/MAP.md (where everything lives, with file and line numbers), atris/wiki/ (what the team learned, as pages linked with [[page-name]]), atris/logs/ (what was done and checked each day), and atris task (open work, who claimed it, and the proof). Every agent reads the same files before it acts, so several agents and people share one memory, reviewed in git like code.
Mem0, Zep and Letta remember the users of a chat app. Atris remembers the project. Patterns and setup: docs/agent-memory.md. Comparison: https://atris.ai/blog/memory-for-coding-agents
Check what ChatGPT can read about your business
atris aeo scan checks whether ChatGPT, Claude, Perplexity and other answer engines can read your site. No login needed.
npx -y atris aeo scan yoursite.com --vs competitor.com
You get a score out of 100, each gap with its fix, a side-by-side score against a competitor, and a draft llms.txt you can publish. To see what ChatGPT actually says when buyers ask about your business, run the free check at https://atris.ai/aeo/check. Guide to both checks and the tools that run them: docs/answer-engine-checks.md.
Quick Start
npm install -g atris
cd your-project
atris init && atris
For Coding Agents
Read FOR_AGENTS.md first: why agents adopt Atris, 60-second start, spread-it instructions.
If you're an agent, run:
Then read the workspace's atris/atris.md and follow it exactly. atris.md is the source of truth.
What Atris Gives You
- An AI workspace on top of any repo
- Persistent AI computers for scoped jobs
- A local computer card that makes each workspace inspectable
- A strict
plan -> do -> review loop
- Daily logs, task tracking, feature packs, and project memory
- Skills, team members, integrations, and cloud sync when you need them
- A plain answer to "what did my AI actually do?": run
atris recap, or atris recap --share for a receipt-backed summary you can paste anywhere
Critical Files
| File | Purpose |
|---|
atris/atris.md | Main instructions for agents working in this repo |
atris/MAP.md | Navigation index with file:line refs |
atris/TODO.md | Human-readable task view |
atris/logs/YYYY/YYYY-MM-DD.md | Daily log, inbox, notes, completions |
atris/features/ | Feature packs with idea.md, build.md, validate.md |
atris/skills/ | Reusable skills for agents |
atris/team/ | Team member specs and local context |
atris/wiki/ | Repo memory, briefs, concepts, and wiki health |
atris/
├── atris.md
├── MAP.md
├── TODO.md
├── logs/
├── features/
├── skills/
├── team/
└── wiki/
Install
npm install -g atris
atris --version
Requires Node.js 18+.
atris task uses built-in SQLite and requires Node.js 22+.
If you want Atris cloud workspaces, businesses, or integrations, run atris setup after install.
How To Run Atris
cd your-project
atris init
atris
atris init scaffolds the workspace, including atris/wiki/. atris loads context and hands the workflow off to atris/atris.md.
If you're still shaping the idea, use atris brainstorm. If you want Atris to finish one job and prove it worked, use atris run "<objective>". If you want Atris to keep picking the next mission or member until you say stop, use atris autopilot. If you want project memory checked for stale pages and missing context, use atris loop. atris activate surfaces wiki state from atris/wiki/STATUS.md when it exists.
Core loop: plan -> do -> review
Integrates with any agent.
Publish a static site
Upload any built static site to an Atris-hosted subdomain. See
docs/site-publish.md for flags, limits, profiles, and
troubleshooting.
atris site publish dist --slug my-site --spa
atris site publish . --slug my-next-app --profile app --build
Bring people together with the Host
The Host is a team member that gets to know everyone in a room (a company, a club, your own network) through one fun question at a time, keeps a card for each person, and introduces two people only when both say yes. Then it offers to find a time and checks that it happened. It never reads anyone's messages and keeps no scores.
atris member install host
atris host setup --room personal
atris host import people.jsonl
atris host room
Put members to work
Members are helpers that work in their own copy of the repo. Three words cover it.
Run hands over one job. Atris picks the right member and starts it.
atris member run "fix one failing test"
Wake keeps members working, each one pulling the next task that fits. Give a number, or name the members you want.
atris member wake 5
atris member wake navigator executor
Wake only picks jobs that have a test to pass, so members know when they are done. Add one like this:
atris task new "fix the flaky login test" --verify "node --test test/login.test.js"
Each job gets the time limit set for building in your roster, plus a few minutes for review. To see or change it:
Naming just one member runs one step of work. Add --keep to keep that one member working.
atris member wake executor --keep
Sleep stops the members that are awake.
To watch them, open the board, or follow one member's steps as they happen.
atris member board --watch
tail -f ~/.atris/progress/executor.log
When a member finishes, ask for its result: what it changed, its receipt, and one number. Each member's copy of the repo stays until its work lands, then tidy puts it away.
atris member result executor
atris member tidy --dry-run
The older words still work as aliases: dispatch and go do what run does, and up does what wake does.
How member work gets checked and lands
Every member job ends with its own check, a command that must pass. Then a second engine reviews the change, and it is never the engine that did the work.
Work lands on its own only after you switch self-landing on, once. From then on Atris pushes the branch, opens a pull request, and turns on GitHub auto-merge, so GitHub merges it when the branch is up to date and CI passes.
Anything else keeps the work on its branch, with one plain line saying why. That covers a failed check or review, unsaved changes, a clash with the main branch, and work that touches a protected area like payments, deploys, or credentials.
A job that fails three times in a row stops and waits for a person.
Chat With Atris 2
ax is the Atris 2 chat and coding-agent CLI. It uses the hosted Atris cloud by default, streams text, shows tool activity, and keeps fresh installs away from local setup.
cd your-project
ax chat ax --fast
ax --fast "hello"
ax --pro "help me plan this task"
ax --max "reason through this hard decision"
ax --doctor
Use Fast for quick chat and low-latency Atris loops. Use Pro for deeper tool loops. Use Max for the hardest jobs; it runs the highest-reasoning model and takes longer per turn. Advanced local workspace mode is opt-in with --local plus AX_BACKEND_URL.
Play AgentXP
AgentXP is the proof-backed game loop for getting better with agents. Start it
inside any project folder:
npm exec --yes --package atris@latest -- atris play
The first run creates a local starter mission if one does not exist. The loop is:
start -> proof -> accept -> login -> sync
The player path:
atris play
atris task claim <mission-ref> --as game-manager
atris task ready <mission-ref> --as game-manager --proof "<artifact path + verifier result>"
atris task accept <mission-ref> --as justin --proof "<human review>"
atris xp card --local
atris login
atris xp sync --local
The manager path:
No proof, no AgentXP. atris task done --proof records review/RL context, but
only human atris task accept mints Career XP for the local card or hosted
leaderboard.
Agents That Sign In On Their Own
atris login signs a person in. An agent can sign itself in with an agent
client, so no one has to hand it a login token.
The owner makes the client once. The secret is shown one time, or written to a
private file with --secret-file:
atris agent-client create --name nightly-bot --scopes mcp:workspace:read,x-search --secret-file ./bot.secret
atris agent-client list
atris agent-client rotate <client_id>
atris agent-client revoke <client_id>
The agent signs in with the id and the secret file. Its short token renews on
its own; the secret and token are never printed:
atris login --client-id atc_... --client-secret-file ./bot.secret
Or set ATRIS_CLIENT_ID plus ATRIS_CLIENT_SECRET_FILE (or
ATRIS_CLIENT_SECRET) and run atris login once. It reads them, no flags
needed, and opens no browser. After that every command renews the token on
its own. ATRIS_TOKEN, when set, still wins.
Business Owners
If you want a shared owner for a company, lab, collective, community, artist, team, or project, use the business command instead of raw atris init.
atris business init "Example Co" --owner-email operator@example.com
cd ~/arena/atris-business/example-co
atris business onboard --website https://example.com --contact "Example Owner" --note "sample customer workspace"
atris align --fix
That creates the shared owner, creates its first/default computer, writes .atris/business.json, initializes .atris/state/ for events and run history, and scaffolds the local atris/ workspace under ~/arena/atris-business/<slug>/ with starter roles, a default recap template, and an initial task queue in atris/TODO.md.
If you do not have a neat source pack yet, atris business onboard is the easiest intake step: give it a website, a named human, a few notes, or run it in a folder with loose files. Atris turns that into raw intake, a starter brief, a first workflow, a safe next action, and a short operator brief.
Use the owner's language when you talk about it:
Your business runs on Atris.
Your lab runs on Atris.
Your collective runs on Atris.
You can also use bare input:
atris business onboard https://example.com "founder-led b2b ops" ./notes.md
If you already have a folder full of source material, run it from there with atris business init "Example Co" --here.
When the first recap is done, record it into the RL state logs:
atris business record atris/reports/2026-04-12-operator-recap.md --outcome mixed --metric "operator speed"
Core Commands
| Command | Purpose |
|---|
atris | Load context and start |
atris init | Scaffold an Atris workspace |
atris brainstorm | Explore before planning |
atris plan | Create the plan/spec |
atris do | Execute work |
atris review | Validate work and capture learnings |
atris run | One bounded mission pursuit: start or resume, tick, complete (--legacy for the old plan/do/review chain) |
atris run logs | Browse glass run logs (phase reasoning persisted to disk) |
atris run prune-logs | Prune old run logs, keeping only the most recent N |
atris run search | Search phase reasoning across all run logs by keyword |
atris run stats | Show run log stats: phase counts, avg durations |
atris run export | Export all run logs as a JSON bundle |
atris run diff | Compare two run logs side by side |
atris autopilot | Keep the workspace moving: mission/member legs until atris autopilot stop (--legacy for the old approval loop) |
atris pulse | Install or run the durable overnight heartbeat |
atris log "note" | Add an inbox item to today's journal (log --repl for the editor; wish for headless intake) |
atris now | Show the current operating truth |
atris radar | Show live agents joined with tasks, missions, and worktrees |
atris ctop | Show process-first live agent CPU and memory |
atris status | Show active work and completions |
atris task | Durable local task state and the agent work loop |
atris mission | Durable goal, owner, verifier, tick receipts, and loop state |
atris play | Enter the AgentXP player loop for one proof-backed mission |
atris gm | Enter AgentXP General Manager mode for player missions and review queues |
atris xp | Show the local AgentXP card and sync eligible proof to the hosted leaderboard |
atris codex-goal | Inspect native Codex task goals and keep completed tasks closed |
atris learn | Manage structured learnings |
atris ingest | Stage raw evidence into atris/context/ and compile into atris/wiki/ |
atris loop | Refresh wiki health, stale/orphan signals, and next ingest candidates |
atris wiki | Full wiki namespace: ingest, query, lint, search, log, and loop |
atris receipt | Save evidence from an agent run |
atris channel | The team's shared channels across computers: atris channel <name> reads, atris channel <name> "message" posts, claim and done take a task (--business <slug>, --as <member>) |
atris pay | Pay another Atris account in credits (100 credits = $1); history, show, accept, cancel |
atris trust | See an account's trust and the friends you share; vouch or unvouch for someone |
atris experiments | Run small experiments and compare results |
atris computer card | Show or write the local owner/computer card |
atris release | Draft or publish a version bump, tag, GitHub release, and launch post |
atris slop | Deterministic slop detector: frontend/prose tells, plus slop dead --exports for dead code |
atris clean | Housekeeping: heal MAP refs, archive journals, report stale pages and dead code |
Doctor
atris doctor [--json] checks Node, task support, authentication, and workspace readiness.
Document health
atris doc-health [--json] [--questions <path>] scores document size, map coverage, lookup hops, and feature/member freshness from 0 to 100, and lists similar feature names. Tokens are estimated as characters divided by four. Boot load earns all 20 points at or below 80,000 characters, falls linearly to zero at 200,000, and stays at zero above that. Questions default to atris/doc-health/questions.jsonl, one {"q":"where is the CLI router","expect":"bin/atris.js"} per line. Custom paths are relative to the workspace root. Missing questions skip lookup scoring and contribute zero of its 30 points. Member age uses the newest log file's modification time.
Built-In Systems
atris learn stores structured project memory in atris/learnings.jsonl
atris wiki keeps repo memory in atris/wiki/ by default, with --cloud when you want the remote workspace path
atris ingest now stages local source packs under atris/context/_ingest/, writes a manifest receipt, and refreshes atris/wiki/STATUS.md plus log.md
atris wiki --private stores local-only sensitive notes under .atris/presidio/
atris loop refreshes atris/wiki/STATUS.md and atris/wiki/log.md, flags stale/orphan pages, and suggests the next ingest
atris activate loads the current wiki status so the next session starts with project memory, not just tasks
atris member keeps team-member identity and learning local-first: MEMBER.md is the role contract, goals.json is the machine-readable goal/experiment state, goals.md is the human readout, and logs/YYYY-MM-DD.md records what happened. Use atris member goal, tick, status, block, and review <name> <experiment-id> --accept --proof "..." --value 1..5 to test whether a member is making useful progress or needs the operator/orchestrator.
atris codex-goal is the read-only bridge for native Codex /goal: status inspects the live goal database, while the legacy reset command now refuses every mutation. Active tasks continue in place; completed tasks retain their final goal, and new work or recurring monitors use a new dedicated Codex task.
atris task keeps durable local task state and append-only events for agents; atris/TODO.md is just a regenerated readable board. Run the loop with atris task new, delegate "..." --to <owner>, next, say, and ready <id> --proof "..."; human approval is atris task accept <id> (moves to Done, awards Career XP) or revise <id> --note "...". Final task transitions also append the general daily log and, when a real atris/team/<member>/MEMBER.md matches the task owner, that member's daily log. Add --json for headless agents, atris task serve for the local board, and atris task show <ref> / events --all for the full ledger. The board prints a private link with a fresh token on each run and listens only on 127.0.0.1. Open that full link in the browser; API clients send its token in x-atris-token and use application/json for changes. Browser changes must come from the board's own origin. Commands accept semantic refs (OBL-18), full IDs, or any unique prefix. In cloud business workspaces, Supabase tasks is the source of truth and Swarlo the live claim layer.
atris mission is the durable autonomy layer: start with an owner, verifier, runner, and stop condition; record bounded work with mission tick; close with mission complete only after proof. Runners include manual, claude, atris2, and codex_goal.
atris pulse is the OS-cron heartbeat for overnight self-improvement. Use atris pulse status, tick, run, install --cadence "<cron>" --days 7 --verify "npm test" --model claude-opus-5-5, and uninstall.
atris experiments runs small test packs in atris/experiments/
atris pull and atris push sync cloud workspaces and journals
atris live keeps a business brain fresh by checking/fixing the workspace, pushing local state, pulling cloud state, and pushing again after local changes go quiet
Agent page (beta)
Agent page is one small browser file, in beta. An AI agent reads the controls on a page by their names, and presses them, without screenshots. It works on the page a person already has when that page is built from normal form controls and ARIA roles. It has no dependencies. The limits are listed below.
The marks are a guardrail for well-behaved agents, not security. A page must still enforce its real rules on its server. Do not put real secrets in a page and rely on a mark to hide them: any software running in that page can read them.
Load it with a plain script tag, or require it in Node and bundlers:
<script src="node_modules/atris/lib/agent-page.js"></script>
const agent = AgentPage.createAgent({
allow: (request) => window.confirm(`Let the agent ${request.action} "${request.label}"?`),
timeout: 5 * 60 * 1000,
});
agent.snapshot();
await agent.act({ steps: [{ label: 'Project name', action: 'set', args: { value: 'Atlas' } }] });
await agent.query({ text: 'Role', kind: 'select' });
Three calls do the work.
snapshot() lists the controls. Each one has a ref, a kind, a label, its value or checked or open state, and the actions it accepts right now. The list is capped at 100. A disabled or read-only control says so and accepts no actions.
act({ steps }) runs steps in order and stops at the first failure. Each step names a control by ref (from a snapshot) or by label, and gives an action: click, set, check, or uncheck. Each result shows the changes, the current value, and any error.
After a change, the file finds the control again in the live page (the same element, or else the one visible control with the same kind and name) and reads its value there. A typed value, a tick, or a choice that did not stick reads as a failure, and so does a press on something that should open or close but did not. A plain press on a button or link says ok once it is delivered; the changes list shows what the page did with it. A disabled choice is refused.
A select or listbox that holds more than one choice shows every chosen value as a list. To change it, send set the full list of choices to keep, for example args: { value: ['Email', 'Chat'] }. A single value is refused when it would clear other choices, so nothing is cleared unless the agent names the whole list. Hidden choices in such a list are left as they are.
query({ text, within, kind }) finds controls and lists the choices of a select or listbox. The list is capped at 50. With text, only the matching choices are listed, so a choice past the cap can still be found.
Three marks on the page set the rules, with the allow callback.
The allow callback decides, and it may return a promise. If you pass one, it runs before every change the agent makes, so the person sees each click, change, or tick first. It gets the control's label, the action and its args, whether the step is marked for confirmation, and the page's own confirm sentences in reasons. If the callback throws, the step fails with a fixed sentence; the error's own message is never passed to the agent.
Before asking, the file lists the elements the step involves and the confirm marks around them. It lists them again after the answer and refuses the step if anything differs: a mark added, removed or reworded, or the control moved under a different mark. It reads the marks again before each part of the change too (after focus, between events, and before each choice it clicks), and stops with the rest undone if a new mark appears.
One approval at a time per page. While a step waits for the person, every other agent on the page is refused, so one agent cannot press another's confirm button. A host's own confirm dialog should still carry data-agent-hidden, so no agent can see it at all.
data-agent-confirm marks a control that needs a person. A step involves the control, its labels (wrapping and for), the form it belongs to (also through the form attribute), every field of that form when the press submits or resets it, each choice the step picks, and each choice or radio the step clears, hidden ones included, with their labels. A mark on any of these, or around them, counts. Without an allow callback, that step fails and tells the agent to ask the person.
data-agent-hidden leaves a region out entirely. On a single choice, it leaves out that choice. A choice hidden by the hidden attribute or by css is left out the same way.
data-agent-secret hides the value and all state of the marked control and everything inside it: value, checked, selected, open, and its choices all read [redacted], and its actions do not change with its state. Its name stays visible, and comes only from a label element, aria-label, or aria-labelledby; placeholder, title, alt and the control's own text are never used. To hide the name too, use data-agent-hidden. Other text inside a secret mark is not read. The agent can still type into a secret field or pick a secret choice it already knows. Password, card, and one-time-code fields count as secret without a mark.
A mark covers the element and everything inside it. Labels, values, changes and error sentences are built when the result is returned, with the marks as they are then: a value that became secret reads [redacted] on both sides of a change, and a control that is gone or hidden is reported by its ref only.
A page can tell when the agent is acting. While an act step runs, document.documentElement has data-agent-acting="true". Every event the agent fires carries event.agentAct = { ref, action }.
What it cannot do:
It reads the main page only. Controls inside iframes and shadow roots are not listed, so the agent cannot reach them. Snapshot and query add an entry of kind unread that counts what was skipped, so the agent knows. A custom element may hold a closed shadow root, so each one is counted as possibly unread.
These limits are accepted. A closed shadow root on an ordinary element, or one made before this file loads, cannot be detected at all. Frames from other sites cannot be read. The file trusts the page it runs in: page scripts can always change anything, including the marks.
A page whose own scripts add, remove or move marks, or swap controls, in the middle of an event is handled only as far as the checks above go: marks read again before each part of a change, and the control found again after it. Beyond that it is out of scope.
Editable text regions (contenteditable) are not typed into. Plain ones are counted as unread. One with a textbox role is listed with its text and no actions.
The tests here drive real Chrome. The controlled-input test imitates React with a hand-built value tracker; a check against real React is planned and not done yet.
For agents that speak tools, agentTools lists three tools (page_snapshot, page_act, page_query) with JSON schemas. registerTools(agent, { allow }) registers them on document.modelContext when the page offers it, and does nothing otherwise.
The mark names and the snapshot, act, and query shape come from Vercel's UI SDK agent bridge. The code here is our own.
Engines
An engine is the intelligence that builds a mission. Every installed headless coding CLI is a swappable worker behind one contract: a bounded prompt goes in, verified proof comes out, and engines never self-certify their own work. The house default is atris-fast (ax); claude, codex, cursor, devin, grok, and others join when their CLIs are installed.
Who does each job
ROSTER.md says which tool and model does each job. Edit it any time; every project, team member, and new chat follows it.
## build
- claude code, model: opus 5.5
## review
- codex, model: gpt-6-astra, effort: medium, max: 20 min, prep: search
- claude code, model: opus 5.5
The first ready worker leads and the rest are backups. ~/.atris/ROSTER.md covers every project on the machine; atris/ROSTER.md overrides it for one project. A worker that stalls sits out 30 minutes and the next one takes over.
The roster warns when a worker runs an old model and a newer one in the same family is on this machine, and prints the command that moves it. It also warns when a pinned model has no proof in 30 days. A landed run is proof; so is verified 2026-09-30 on the line after a quick smoke test.
atris engine roster
atris engine assign build claude --model "opus 5.5"
atris engine roster --available
atris team
atris team --record
Show every installed engine and the current default:
Flip the default for this workspace (persists to .atris/engine.json):
atris engine cursor
atris engine reset
Ride a different engine for one run only, without changing the default. --engine <name> works on mission run, autopilot, and run. Per-run flags and ATRIS_RUNNER_PROFILE always beat the saved file.
atris mission run "fix the flaky login test" --engine codex
Fleet Flight
atris mission run --fleet staffs every idle installed engine on the board's claimable safe-lane tasks: one mission per task, one worktree per engine. Builds run in parallel; arrivals land serially with rebase-before-ship, so a rebase conflict pauses that landing (never auto-resolved) and keeps its worktree instead of orphaning work. Each flight writes a receipt to atris/runs/fleet-<stamp>.json.
atris mission run --fleet --slots 3
atris mission run --fleet --dry-run
Each job goes to the builder its real results favor. The fleet counts what each engine landed and missed since Nov 1, on jobs like this one and on the rest (the same plays atris rating shows), gives every engine a fair chance, and prints the pick under the job with the numbers behind it. A landing counts only when its commit is on origin's default branch (the branch origin/HEAD names, never a guess), is no older than the claim it lands (five minutes of clock skew allowed), and no other claim recorded the same commit first: a commit counts once, for the first claim that recorded it, and not again for a later claim on the same job (reopened and claimed again, by the same engine or another) or for another job. That is judged on every landing record in the task store, in every workspace that shares it and every later review pass. A landing pays the engine that finished the job: when a stalled or dead engine's job is finished by another, the engine that finished it is credited, and a landing whose hand-in names several engines without saying which finished counts for none. The folder is judged in one order. With no git, no origin, or origin/HEAD not set (git remote set-head origin --auto sets it), no landing can be checked, so jobs are dealt in turn and the flight says why in one line. With an engine's landing on record, the flight fetches origin's default branch once before it picks (so a merge from a minute ago counts); if that fetch fails, jobs are dealt in turn with one line, and if it works the landings are checked. With only misses on record, they steer on their own, with no fetch. Everything git does before a pick shares one five-second budget and never prompts: its input is closed, GIT_TERMINAL_PROMPT is 0, and ssh runs in batch mode unless you set your own ssh command, which is left as it is and mentioned in one line. A play counts toward the kind of job it was when the engine took it: the claim, whether made by the fleet or by task do, saves the job's title and tags, so tagging or renaming a finished job later changes nothing, and a claim made before that was saved counts at half weight. New engines get tried; engines that keep missing get fewer jobs. An engine that is out of credit, cooling, or one the Claude spend guard would stop (a paid model nobody approved, or a pay-per-use key) is never picked, and the flight says which. The guard's own hook (atris spend guard install) stays the real guard on every unattended Claude run; this check only avoids picking an engine the hook would refuse, so where it is unsure it leaves Claude out. The fleet's own launch is judged by the --model the fleet itself passes, read through the ANTHROPIC_DEFAULT_*_MODEL remaps in the environment the worker will really get (a sealed one-lap worker gets none); with no --model, by ANTHROPIC_MODEL in that same environment, else the paid default. No Claude settings file is read, from any home. A custom launch template (ATRIS_RUNNER_COMMAND_TEMPLATE) is never read for its model, because a model in it can be a flag, a word of the prompt, another command's, or one the shell works out only when it runs. It counts as a model nobody can tell, which counts as paid, so Claude is left out with one line saying so, unless ATRIS_CLAUDE_SPEND_OK approves default; --engine claude still uses it. Your own choices come first and are never picked for, with results or without: --engine <name> sends every job to one engine, a job that names its engine gets it, and a job assigned to a member goes to that member's own engine, the same one atris member uses. A choice the flight cannot give is never given to someone else: a job that names an engine this flight does not have, or whose member's engine it does not have, is left out of the flight, with one line naming the job and why. It is taken out before the flight is staffed, so it takes no slot and holds no file against another job. The pick goes on who a job is assigned to and the engine it names, as the task list showed them; the claim asks the store for both again, inside its own transaction, and a job that now says something else is not claimed for the old choice. It is left for the next flight with one line, and the task list is drawn again so that flight reads the job as it is now. An engine list you name is never filtered, on the first pick or on a retry. If an engine dies mid-job, the next one is picked the same way; if it stalls and the roster lists a backup order, the roster's order stands. --pick turns (or ATRIS_PICK=turns) deals engines in turn, as before.
Verifiable Feedback Loop
Under the hood, Atris can keep score on real repo work.
- Tasks can carry a
Verify: command, so work can end on a deterministic check instead of pure prose.
- Plan, do, and review now carry the same Confidence Gate: find every plausible loophole, patch it with source/proof/verifier/owner/rollback, or name it as residual risk before advancing.
atris autopilot can run that check after review and record the result in the journal.
- Future task picks can use recent results, so Atris learns from repo-local history without claiming model retraining.
Which AI Can Do Your Work
atris scorecard measures which AI can do your team's work, on your own repo.
atris scorecard
atris scorecard run
atris scorecard run --models haiku,opus
- A past change becomes a task when its tests fail before it and pass after it. The ask is what your team wrote for that change.
- The ask also names what the change's tests call that the change added or changed: a new file, a function with its arguments, an export, found with no AI. A good ticket names these; a commit message rarely does, and no model can guess a name nobody wrote down.
- Each try starts from the change's parent in a throwaway copy with no remote. The change's own tests stay hidden until they grade the try.
- Every try runs on your Claude plan login. Pay-per-use keys are removed, a try stops the moment Claude Code reports any other billing source, and nothing starts with extra usage on unless you pass
--allow-extra-usage.
- The scorecard shows how many changes each model did, how long it took, what the same runs would cost on pay-per-use, and which changes no model could do from what was written.
- Tests that need their own command or settings go in
.atris/scorecard.json.
The Atris Rating
atris rating gives every AI teammate one number, 25 to 99, for how much real work it finishes.
atris rating
atris rating opus
atris rating <member>
- 99 means 99 of every 100 finished. Each halving of misses is worth about 6 points, so the top is the hardest climb: 3 in a row rates 76, and it takes about 60 in a row to rate 99.
- Engines are rated in the combine: Loop Bench tasks from the atris repos, each graded by the tests that shipped with it. No engine is a 99 yet.
- Each engine's card also rates its Python tasks, its JavaScript tasks, the tasks the Atris prompt was never tuned on, and how rarely a try broke another test, with its median time and pay-per-use price per finished task.
- Members are rated on the same scale from the jobs they landed in your workspace over the last 30 days. Landed means accepted or merged; drops and abandons count against them.
atris gm shows the same numbers.
- When a person merges a builder's pull request, the builder is usually credited within the hour, by an
atris brief, atris rating, atris gm or flight that reaches GitHub; atris task sweep --held does it right away. ATRIS_HELD_SWEEP=off stops the automatic check.
- The session boot shows the build engine's rating, for example
build opus 5.5 (86 ovr). The combine grades building only, so review and search engines show none.
Benchmark Harness
Atris ships one public head-to-head benchmark harness for comparing a pinned
single-model baseline against a coordinated stack run on the same task brief.
Quickstart:
node bin/atris.js experiments validate endstate-baseline
node bin/atris.js experiments validate endstate-stack
node bin/atris.js experiments run endstate-baseline --dry-run
node bin/atris.js experiments run endstate-stack --dry-run
node bin/atris.js experiments compare endstate
One-command rehearsal:
node bin/atris.js experiments replay endstate
What to inspect:
- receipts land in
atris/experiments/endstate-baseline/artifacts/ and
atris/experiments/endstate-stack/artifacts/
- scores append to each pack's
results.tsv
atris experiments compare endstate prints the latest side-by-side comparison
atris experiments replay endstate runs the full public dry-run rehearsal
- the benchmark contract lives at
atris/features/endstate/contract.md
- the verification log lives at
atris/features/endstate/validate.md
The stack wins Level 1 only if it beats the baseline on total score and does
not lose the reviewed completion category.
Skills
Atris ships a real skill catalog in atris/skills/, not just one workflow file.
Examples:
atris, autopilot, autoresearch, wiki, loop
backend, design, copy-editor, meta, writing
github, email-agent, calendar, drive, slack, notion, slides, x-search, youtube, ramp
apps, create-app, create-member, memory, magic-inbox, improve, skill-improver, flow
atris skill list
atris skill audit [name]
atris skill fix [name]
atris skill create <name>
atris skill link [--all]
For Codex, copy any skill folder into ~/.codex/skills/.
Before a New Version
Run the release-facing checks before bumping:
node bin/atris.js --help
node bin/atris.js pulse --help
node bin/atris.js release --help
node --test
Then dry-run the release:
atris release drafts or publishes from local git history. --dry-run prints the planned bump without committing, tagging, or pushing.
Update
atris upgrade
atris update
Packaged npm installs check for new versions during normal commands and start a background update automatically. Git checkout installs stay manual so linked development copies are not overwritten unexpectedly.
License: MIT | Repo: github.com/atrislabs/atris