Flow Atelier

Workflows and loops, configured in one simple YAML file. The steps are shell commands, AI coding agents, and human approvals. Run them from your terminal or a local visual editor.

Flow Atelier is a workflow runner whose tasks can hand work to Claude Code, Codex, Gemini, opencode, Copilot, Cursor, or any of the ~40 agents in the ACP registry — each driven through your existing login for that tool. No API keys are configured, stored, or proxied by flow-atelier.

There is no SDK to learn and no code to write: a workflow is a plain YAML file you (or an agent) can read end to end and edit. Steps run in parallel where their dependencies allow, loop until output matches or while it keeps matching, retry on failure, branch on what a previous step printed, pause for a typed human answer, and write every run to disk so you can replay it.

name: ci
description: Run the test suite until it passes
tasks:
  - run_until_green:
      description: retry the tests until they pass, up to 5 times
      task: "make test"
      tool: tool:bash
      depends_on: []
      repeat: 5                   # loop this task up to 5 times...
      until: output.match(PASS)   # ...stopping as soon as the output matches

Save that as .atelier/conduits/ci/conduit.yaml and run atelier run ci.

curl -fsSL https://raw.githubusercontent.com/Andesprit/flow-atelier/main/install.sh | bash
atelier init                             # writes a hello-world conduit
atelier run hello --input name=world     # runs it
atelier serve                            # opens the visual editor on :8000

What it looks like

The designer lays a conduit out by dependency depth, so each column is a set of steps that can run at the same time:

Runs stream to the dashboard, including the human-approval gates:

How it compares

	Flow Atelier	n8n / Zapier	Prefect / Dagster / Airflow	GitHub Actions
Runs where	your machine, local-first	hosted / self-hosted server	scheduler + workers	CI runners
AI steps	a first-class task type, using your own agent CLI login	LLM API nodes, you supply keys	you write the client code	you write the client code
Human-in-the-loop	built in, mid-DAG, blocks the run	via external forms/webhooks	not really	manual approval on environments
Workflow format	one simple YAML file per conduit	JSON built in a GUI	Python	YAML
Loops & retries	repeat / until / while / retries on any task	loop and wait nodes in the GUI	Python control flow	matrix builds; no retry-until
State	plain files under .atelier/	a database	a database	opaque to you

Pick Flow Atelier when the work is a repeatable recipe you want AI agents to execute on your own machine, with you in the loop at the points that matter. Pick the others when you need multi-tenant hosting, distributed workers, or enterprise scheduling — flow-atelier deliberately does none of that.

Note on scope. flow-atelier runs shell commands and AI agents on the machine it is installed on. It is a local developer tool, not a hosted multi-user service. See Security before exposing it on a network.

Why processes, not agents

Flow Atelier is built on a simple premise: the world doesn't run on people, it runs on processes that people execute. Ask someone to make anything, for example: the best yogurt in the world, what a person will do is that they'll research, experiment, and produce something average. The same is true of an AI agent. But give that person a clear, step-by-step recipe — and keep refining it over time — and you can reach the best yogurt in the world. That's what Flow Atelier does. We help you build simple, repeatable instructions for getting something done the same way every time, then let anyone improve on them. It mirrors how real businesses actually work. Coca-Cola and Pepsi don't differ because their managers, lawyers, or engineers are fundamentally different people — they differ because their processes are different. Flow Atelier gives you the tool to design and refine those processes. The difference: instead of people executing them, AI agents do the work.

The words we use

A few terms show up everywhere in this document:

• Conduit — a recipe. An ordered set of steps written in a single YAML file.
• Task — one step in that recipe. A task runs a shell command, asks a person a question, calls an AI tool, or runs another conduit.
• Flow — one run of a conduit. Every run is saved to disk, so you can always go back and see exactly what happened.
• Harness — an AI coding tool (Claude Code, Codex, opencode, Copilot, Cursor) that a task can hand work to.
• HITL — "human in the loop": a step that pauses to ask a person a typed question, then continues with the answer.

What can you build with it?

Any pipeline that can be described as an ordered sequence or graph of steps. The two examples further down — a one-line greeter and a deploy pipeline with a human approval gate — illustrate two possible shapes; they are not prescriptive templates.

A non-exhaustive list of things people have built:

• Chatbots and AI agents that chain multiple turns of conversation with retries, branches, and fallbacks.
• Multi-AI pipelines in which one assistant drafts a specification, a second produces a plan, and a third executes it — or in which two assistants review the same change and a third synthesizes their feedback.
• CI/CD-style pipelines that clone a repository, run tests with retry, request an AI code review, ask a human for confirmation, and then deploy or roll back based on the result.
• Scheduled jobs such as daily reports, weekly syncs, or one-shot reminders at a specific time.
• Polling loops with retry and backoff that call an endpoint until it returns a success status or a rate limit lifts.
• Human-in-the-loop automations — flows that pause to ask the operator a typed question and resume with the answer.
• Reusable building blocks — one conduit invoking another, so a deploy conduit written once can be called from many higher-level pipelines.
• Pure-shell automation with no AI at all — flow-atelier works as a general-purpose task runner in this mode.

If a task can be described as a sequence or graph of steps, it can be written as a conduit.

How a conduit runs

You write a conduit YAML file and put it in .atelier/conduits/<name>/conduit.yaml. When you run atelier run <name>, flow-atelier:

1. Reads the YAML.
2. Looks at each task's depends_on list to figure out which tasks can start now and which have to wait.
3. Starts every ready task at the same time, up to a configurable limit (max_concurrency).
4. For each task, picks an executor based on the tool: field:

Tool	What it runs
tool:bash	a shell command
tool:hitl	prompts a human on the terminal for one or more named answers
tool:conduit	another conduit, as a nested run
harness:claude-code	Claude Code (via the ACP adapter)
harness:codex	OpenAI Codex (via the ACP adapter)
harness:opencode	opencode
harness:copilot	GitHub Copilot CLI
harness:cursor	Cursor CLI

5. Saves everything to disk under .atelier/flows/<flow_id>/ — what ran, what each task printed, whether it succeeded, when it finished.

Every AI harness uses that tool's own login that lives on your machine. flow-atelier never sees, stores, or proxies any credentials.

Install

One-command install (no Python needed)

The quickest way. The script downloads a prebuilt atelier binary into ~/.atelier/bin, verifies its SHA-256 checksum against the published release, and adds it to your PATH. It is safe to re-run to upgrade.

Once a day the binary checks GitHub for a newer release and, if one exists, says so on stderr. Nothing is installed until you run atelier self-update. Set ATELIER_NO_UPDATE_CHECK=1 to silence the check.

macOS (Apple Silicon) / Linux:

curl -fsSL https://raw.githubusercontent.com/LGuillermoAngaritaG/flow-atelier/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/LGuillermoAngaritaG/flow-atelier/main/install.ps1 | iex

Prebuilt binaries are published for Linux x86_64, macOS arm64 (Apple Silicon), and Windows x86_64. Intel Macs are not supported (no Rosetta fallback exists for an arm64 binary). Open a new terminal after installing so the updated PATH takes effect.

Install with uv (for Python users)

If you have Python 3.13+ and uv, you can install from PyPI instead:

uv tool install flow-atelier
uv tool upgrade flow-atelier      # upgrade later
uv tool uninstall flow-atelier    # remove

Either way, you end up with an atelier command on your PATH.

Optional: AI harnesses

You only need the AI tools you actually plan to use. If you never use AI in your conduits, you can skip this entire section.

flow-atelier does not install agents and does not manage their logins. You install the agent you want and log into it with its own CLI; then you point flow-atelier at its command, either by name or by argv. There is no bundled installer, no download manager, and no credential handling here.

What flow-atelier does do is run the command you selected, exactly as that agent documents it. For agents distributed through npx or uvx, the documented command fetches the package on first use — that is the agent's own distribution mechanism doing its normal thing, the same as running the command yourself in a shell. Agents distributed as a binary are never downloaded; you install those, and flow-atelier runs what it finds on PATH.

An AI task names its agent and nothing else:

- review:
    description: review the diff
    task: "review the working tree and list any bugs"
    tool: harness:gemini
    depends_on: []

The names come from the ACP registry, a snapshot of which ships with flow-atelier. To see what you can type and what already works on your machine:

atelier list harnesses           # every agent, and whether it runs here
atelier list harnesses --ready   # just the ones you can use right now
atelier harness sync           # refresh the list from the ACP registry

Roughly 40 agents are listed, including harness:claude-code, harness:codex, harness:gemini, harness:copilot, harness:cursor, harness:opencode, harness:qwen-code, harness:goose and harness:amp-acp. A name is only the launch command that agent documents; the via column says how it starts:

• npx / uvx — the agent's own package manager fetches it on first run, at the version the registry pins. Needs Node.js or uv on PATH.
• binary — you install the agent's CLI, and flow-atelier runs it from PATH. atelier list harnesses names the missing binary when it isn't there.

Either way, logging in is yours to do, with that agent's own CLI.

Picking a model

A third segment names the model, spelled the way the agent lists it:

    tool: harness:codex:gpt-5.1-codex
    tool: harness:claude-code:claude-sonnet-4-5

flow-atelier selects it on the ACP session before the first prompt. A model the agent does not offer fails the task at once and prints the models it does offer. atelier harness check <name> lists them too, as models:. Without the suffix the agent runs on its own default.

Picking a reasoning effort

A fourth segment names how hard that model should think, again spelled the way the agent lists it:

    tool: harness:codex:gpt-5.6-sol:high
    tool: harness:claude-code:opus[1m]:xhigh

Effort only comes after a model, because which efforts exist is a property of the model: the agent is asked for the model first, and the efforts it offers for that model are what the value is checked against. An effort that model does not offer fails the task before the prompt is sent, and prints the ones it does offer.

Leave the segment off and the model keeps whatever effort it defaults to — flow-atelier sends nothing, so your own agent configuration stands.

To see the choices for a given model, no prompt and no tokens:

atelier harness check codex             # efforts for its default model
atelier harness check codex:gpt-5.6-sol # efforts for that model

Checking a harness before you use it

atelier harness check gemini
atelier harness check codex:gpt-5.6-sol
atelier harness check --cmd "/opt/my-agent --acp"

This starts the agent, completes the ACP handshake, opens a session and stops. No prompt is sent, so it costs no tokens. A :<model> or :<model>:<effort> suffix is resolved the way a run resolves it, so what the check reports is what a task naming that same tool would get. It reports one of:

• ok — with the agent's name and version, the ACP version, the session modes it offers, and the models and reasoning efforts it offers for the model in force.
• not found on PATH — install the agent yourself, then re-check.
• started but did not speak ACP — usually the wrong entry point; many CLIs need an --acp flag.
• could not open a session — usually not logged in. The check lists the auth methods the agent advertises, and you log in with that agent's own CLI.
• not usable — model/effort not offered — the session opened fine; the suffix named something the agent does not have. The message lists what it does.

Failures exit non-zero and include the tail of the agent's own stderr, which is where a failing agent explains itself.

For an agent the registry doesn't list — something private, a fork, a local build — give flow-atelier its command and it becomes a first-class harness:

ATELIER_HARNESSES='{"mine":["/opt/my-agent","--acp"]}'   # tool: harness:mine

To pin one of claude-code, codex, opencode, copilot or cursor to a specific argv, the matching ATELIER_*_LAUNCH_CMD variable still overrides the registry (see .env.example).

Quickstart

atelier init                                # creates .atelier/conduits/hello/
atelier run hello --input name=world        # runs it
atelier status latest                       # shows progress of the newest run
atelier list flows --conduit hello          # lists previous runs

atelier init writes a one-line hello conduit that only runs a shell command, so this works end-to-end before you install any AI tool.

Ready for a real one? Your first workflow: run it, break it, recover it is a 5-minute Bash-only exercise that builds a three-step conduit, fails it on purpose, diagnoses it from saved history, and shows what --resume keeps that --again redoes.

Your first AI workflow: review what you staged

atelier create --template code-review writes a two-step conduit that captures the patch you have staged and hands it to Claude Code for a review. Nothing runs at creation time — you get an ordinary YAML file to read, run, and edit.

atelier harness check claude-code   # confirms the agent starts and you're logged in
atelier create my-review --template code-review
atelier check my-review             # validates it and confirms the harness is usable
atelier plan my-review              # prints the two waves, runs nothing
git add -p                          # stage the changes you want reviewed
atelier run my-review
atelier outputs latest --task review

You need a Git repository, bash, and Claude Code installed and logged in (see Optional: AI harnesses for its launcher prerequisites). The review covers only what is staged — git diff --cached. Unstaged edits and untracked files are left out until you git add them, and an empty index skips the review step instead of calling the agent.

The workflow only reads your repository: it never stages, commits, or edits anything for you. The prompt tells the agent to analyse rather than act, which is an instruction to the agent, not a sandbox around it.

If a run fails, atelier status latest and atelier logs latest say which step broke and what it printed. latest means the most recently started run in this project; pass the printed flow id instead when several runs overlap. To change the prompt, the harness, or the diff range, edit .atelier/conduits/my-review/conduit.yaml — it is a normal conduit, and atelier show my-review prints the exact prompt it will send.

Reading a conduit before you run it

Conduits arrive from atelier init, atelier create, a teammate's repository, or an installed package. atelier show prints one without running it — no task starts, no agent is launched, no run is recorded.

atelier list conduits --json          # what is installed here
atelier show hello                    # the exact YAML that would run
atelier show hello --json             # the same thing, normalized, for tools
atelier check hello                   # is it valid, and is its agent usable?
atelier run hello --input name=world  # run it

atelier show hello writes the file's own text to stdout — comments, templates and multiline prompts exactly as written — and the source and path to stderr, so atelier show hello > copy.yaml gives you a clean copy. A project conduit shadows a global one of the same name here just as it does at run time.

--json answers the question a script or a coding agent actually has: how do I call this? (conduit is abbreviated below — the real output carries the whole definition.)

{
  "source": "project",
  "path": "/home/you/project/.atelier/conduits/hello/conduit.yaml",
  "conduit": { "name": "hello", "inputs": { "name": { "description": "Who to greet", "default": null } }, "tasks": [] },
  "accepted_inputs": ["name"],
  "required_inputs": ["name"]
}

• conduit is the whole definition — every task body, tool, dependency, loop and default — normalized, but with templates left unresolved.
• accepted_inputs is every key the conduit can use, including keys only referenced as {{inputs.x}} in a task and keys forwarded to a nested conduit.
• required_inputs is the subset you must pass: the declared inputs whose default is null. An input with default: "" is optional — an empty string is still a default.

So the --input flags a run needs are one command away:

atelier show hello --json | jq -r '.required_inputs[]'   # jq is optional

These are the conduit's declarations, not a promise that every template resolves or that its agent is installed — atelier check still answers that. show reads whichever copy would run even when that copy is broken, so you can see the mistake; --json refuses to guess and exits non-zero instead.

Passing a file as an input

A brief, a spec, a failing build log — the material a workflow needs is usually already a file. --input-file key=path hands that file's text to one input, so you never paste a document into the command line:

atelier init
printf 'Ada' > name.txt
atelier run hello --input-file name=name.txt
atelier outputs latest

--input-file reads the file as UTF-8 and passes the text through unchanged — every character, every trailing newline, no stripping, no YAML parsing, no template expansion of anything inside it. Relative paths are resolved from wherever you typed the command. Both flags mix freely, so the long thing comes from a file and the short settings stay literal:

atelier run my-review --input-file brief=SPEC.md --input tone=blunt

Use atelier show <conduit> --json to see which keys a conduit accepts. A key can be given by --input or --input-file, never both: a repeated key is a usage error rather than one value silently winning. A missing file, a directory, non-UTF-8 bytes or a malformed pair fails before the run starts, so nothing is recorded. - is rejected too — stdin stays free for the questions a tool:hitl step asks you.

The loaded text is saved with the run like any other input, not a reference to the file. atelier run --again <flow_id> therefore replays the text the run actually used even if you have since edited or deleted the source file; pass --input-file again to feed it a new version. --resume is refused with --input-file, because resuming continues a flow's saved inputs rather than starting a new run.

File contents are ordinary input values: a conduit that drops an input into a tool:bash command interpolates it the same way it interpolates anything else. --input-file is a convenience for passing documents, not a sandbox and not secret storage.

Writing a conduit with your editor's help

atelier schema prints the JSON Schema of a conduit.yaml, generated from the models the version you have installed actually loads. Save it next to your workflows and an editor with a YAML language server will complete the field names and underline the mistakes as you type:

atelier init
atelier schema > .atelier/conduit.schema.json

Then make this the first line of .atelier/conduits/hello/conduit.yaml:

# yaml-language-server: $schema=../../conduit.schema.json

The path is relative to the conduit file, so ../../ lands on .atelier/. You need an editor with the YAML language server for this — the VS Code YAML extension, or the same server through your own LSP client. Nothing is installed for you and nothing is sent anywhere.

Now break something on purpose:

# yaml-language-server: $schema=../../conduit.schema.json
name: hello
description: Say hello
max_concurrency: 0

The editor marks 0 with Value is below the minimum of 1 before you run anything. Delete the line and the mark goes; type re inside a task body and it offers repeat, retries, retry_backoff, until, while and the rest. Both shorthands are covered, so name: Who to greet under inputs: and - greet: with the body indented under it are as valid to the editor as the long forms. Then the usual sequence:

atelier check hello
atelier run hello --input name=world
atelier outputs latest --task greet

A coding agent does not need the file at all — atelier schema on its own is the whole vocabulary, from the version that is installed.

It checks shape, not meaning. A missing tool, a tasks: that is not a list, a max_concurrency: 0 — yes. A depends_on naming a task that does not exist, two tasks with the same name, a loop predicate that will not parse, a template that resolves to nothing, an agent you never installed — no. atelier check <name> still owns all of that, and it is still the thing to run before a real workflow. Regenerate the file after upgrading Atelier; the schema describes the version that wrote it.

A field Atelier does not know is an error

A conduit.yaml may only use the fields Atelier defines. A misspelled one used to be dropped in silence, which is the worst possible outcome: the file checks, plans and runs, but not as the workflow you wrote.

name: typo_demo
description: two tasks that are meant to run in order
tasks:
  - name: prepare
    description: write the sentinel
    task: "printf 'prepared\n' > sentinel.txt"
    tool: tool:bash
  - name: consume
    description: read the sentinel back
    task: "cat sentinel.txt"
    tool: tool:bash
    depend_on: [prepare]      # typo: the field is depends_on

Before, consume loaded with no dependencies at all and raced prepare for a file that did not exist yet. Now:

atelier check typo_demo

typo_demo [project] — FAIL: tasks[1].depend_on: Extra inputs are not permitted

Correct it to depends_on: and the same file checks, plans consume into the second wave, and runs. The API rejects the same fields, so the designer and a coding agent posting JSON get the identical answer.

If you kept your own notes inside a conduit: anything Atelier does not define was already being discarded on load, so it never reached a run — but it is now an error rather than a silent drop. Move it to a YAML comment or a file beside the conduit. And if you exported conduit.schema.json before upgrading, run atelier schema again: the old copy still accepts the typo your editor should now be underlining.

Checking conduits from a script or an agent

atelier check --json answers the same question as atelier check, in a form a program can act on. It writes one array to stdout, one object per conduit it checked, and exits 1 if any of them failed:

[
  {
    "name": "hello",
    "source": "project",
    "path": "/home/you/project/.atelier/conduits/hello/conduit.yaml",
    "ok": true,
    "error": null,
    "required_inputs": ["name"]
  }
]

• path is the file to open to fix the problem — the copy that would actually run, so a broken project conduit is reported instead of the working global one it shadows. It is null only when the path could not be resolved at all.
• ok is the verdict. Branch on it; the wording of error is for a human to read and may change.
• required_inputs is the --input keys a run needs, and only appears when the check passed — a conduit that failed to load has no trustworthy input list, so it is null, never [].

Exit status and stdout carry different information, so read both:

atelier check --json > check-report.json   # exit 1 when a conduit failed
status=$?

python3 - <<'PY'
import json
for row in json.load(open("check-report.json")):
    if not row["ok"]:
        print(row["path"], "->", row["error"])
PY
exit $status

Exit 1 with a report on stdout means "checked everything, some failed". Exit 1 with empty stdout means the check never started — an unknown conduit name, or a store that could not be read — explained on stderr. An empty [] with exit 0 means no conduits are installed here, which is not proof that anything was validated.

With no name it checks every conduit including the global ones in ~/.atelier/conduits/, so a report can name a file outside this project. Results depend on this machine: a conduit that needs an agent CLI you have not installed fails here and passes where it is installed. Readiness means the tool is available, not that it is logged in or that the run will succeed.

So a coding agent can repair a workflow without a human reading the terminal: atelier schema for the vocabulary, write the YAML, atelier check <name> --json, open the path it returns, fix the error, check again, and run it once ok is true. Nothing runs during a check — no task, no agent session, no recorded flow.

Checking a workflow that calls other workflows

A tool:conduit step runs another conduit by name, and that child can call one of its own. Plain atelier check stops at the conduit you named: a child is only loaded once the run reaches that step, so a missing or broken child surfaces after the earlier steps have already done their work. atelier check <name> --recursive follows those calls first.

Build a two-level workflow in a throwaway directory — a shell step that prepares something, then a call to a summary conduit that does not exist yet:

workspace="$(mktemp -d)/composed demo"
mkdir -p "$workspace/.atelier/conduits/report" && cd "$workspace"

cat > .atelier/conduits/report/conduit.yaml <<'YAML'
name: report
description: Prepare a measurement, then hand it to the summary conduit
tasks:
  - name: prepare
    description: record that preparation happened, and measure something
    task: "echo prepared >> preparation.log && echo 42"
    tool: tool:bash
    depends_on: []
  - name: summarise
    description: turn the measurement into a summary
    task: summary
    tool: tool:conduit
    depends_on: [prepare]
    inputs:
      finding: "{{prepare.output}}"
YAML

The parent file itself is fine, so the ordinary check passes and the recursive one does not:

atelier check report              # OK - nothing is wrong with this file
atelier check report --recursive  # FAIL, exit 1 - `summary` is missing

The failure names the step that makes the call and the name it could not resolve:

report [project] — FAIL: report.summarise -> summary — conduit not found

Neither check ran anything: there is no preparation.log and no recorded flow. Write the child, then gate the run on a passing recursive check:

mkdir -p .atelier/conduits/summary
cat > .atelier/conduits/summary/conduit.yaml <<'YAML'
name: summary
description: Write a one-line summary of a finding
inputs:
  finding:
    description: what the caller measured
tasks:
  - name: write
    description: write the summary line
    task: "echo summary of {{inputs.finding}}"
    tool: tool:bash
    depends_on: []
YAML

atelier check report --recursive \
  && atelier run report \
  && flow_id=$(atelier wait latest --timeout 60) \
  && atelier outputs "$flow_id" --task summarise

The last line prints summary of 42 — the child's result, read back from the parent's saved run — and preparation.log holds exactly one line, because the two failed checks never executed a step.

The call's arguments, not only the child's name

A called conduit receives exactly what its calling task forwards under inputs:. It does not inherit the caller's inputs, so that one map is the whole interface between the two files — and --recursive checks it.

Extend the same example: give summary a second input with a default, and let the caller misspell it.

name: summary
description: Write a one-line summary of a finding
inputs:
  finding:
    description: what the caller measured
  tone:
    description: how the summary should read
    default: calm
tasks:
  - name: write
    description: write the summary line
    task: "echo {{inputs.tone}} summary of {{inputs.finding}}"
    tool: tool:bash
    depends_on: []

name: report
description: Prepare a measurement, then hand it to the summary conduit
tasks:
  - name: prepare
    description: record that preparation happened, and measure something
    task: "echo prepared >> preparation.log && echo 42"
    tool: tool:bash
    depends_on: []
  - name: summarise
    description: turn the measurement into a summary
    task: summary
    tool: tool:conduit
    depends_on: [prepare]
    inputs:
      finding: "{{prepare.output}}"
      tonee: direct

The engine drops a key the child cannot use, so this run would have succeeded and written a calm summary — the default — while the author believed they had asked for direct. atelier check report --recursive fails instead, on one line naming the call, the child's file, the bad key and the nearest one it does know:

report [project] — FAIL: report.summarise -> summary (/tmp/demo/.atelier/conduits/summary/conduit.yaml) — task 'summarise' has unknown inputs: ['tonee'] (did you mean 'tone' for 'tonee'?); 'summary' accepts inputs: ['finding', 'tone']

Rename tonee to tone and the check passes. Delete the finding: line from the corrected file and it fails again, because finding has no default and nothing else can supply it:

report [project] — FAIL: report.summarise -> summary (...) — task 'summarise' supplies no value for required inputs: ['finding']; add them under the task's own 'inputs:' map, which is all a called conduit receives

Both used to be found only by running the workflow — the second after prepare had already appended to preparation.log, the first not at all.

What the flag does and does not tell you:

• The report still has one row per conduit you selected, with the same --json keys. A nested failure sets that row's ok to false and puts the calling chain, the child's file and the real diagnostic into its error.
• required_inputs stays the root's inputs. summary declares finding with no default, but the calling task supplies it, so atelier run report needs no --input.
• Binding names are checked: a forwarded key the child can neither declare nor reference, and a child input declared with no default that the call leaves out, both fail. Binding values are not. A {{...}} you forward is accepted without being resolved, and a key the child only references — supplied at run time by a loop, a human answer or an upstream output — is neither required nor rejected here.
• Every tool:conduit step is inspected, including one a condition would skip at runtime.
• A target assembled from a template (task: "{{inputs.which}}") cannot be resolved without running the workflow, so --recursive fails and says so. Check that child by its real name instead, or leave the flag off.
• A conduit that calls itself, or a loop between two conduits, is reported as a cycle instead of recursing — as is a chain deeper than the engine's nesting limit.
• Passing still only means "these definitions load and their tools are on this machine". It is not a promise that the run will succeed.

Waiting for a run from another terminal or agent

A run started in another terminal, by the dashboard, or by the scheduler is an ordinary saved flow, so a second session can join it and use its results:

flow_id=$(atelier wait latest --timeout 60) && atelier outputs "$flow_id" --json

atelier wait watches one run's saved progress and turns the outcome into an exit status, so nothing has to poll, re-read status --json, or scrape the terminal:

exit	what it means
0	the run saved completed; stdout holds the resolved flow id and nothing else
1	it failed, was stopped, its runner died, or its progress could not be read
124	the timeout expired while it was still running
130	you pressed Ctrl-C
2	--timeout was not a positive number of seconds

Success prints the id it resolved, and the next command should use that id. Resolving latest a second time can land on a newer run that started while you were waiting.

Waiting only watches. A timeout or a Ctrl-C ends your observation, not the run: the other process keeps going and atelier status <flow_id> still finds it. wait never starts, stops, resumes or edits anything. A run paused on a human gate (tool:hitl) counts as still running, so it times out rather than failing.

Failures name the state and point at atelier status and atelier logs; a run whose runner died also suggests atelier run --resume.

Examples

The two conduits below are illustrative, not prescriptive. A conduit can have one step or fifty, and any combination of shell, AI, and human steps. The samples show one minimal conduit and one larger one to demonstrate the range; the conduits you write will look nothing like them.

A simple conduit (hello)

The one-task conduit that atelier init creates. It runs a single shell command:

name: hello
description: Say hello
inputs:
  name: Who to greet
tasks:
  - greet:
      description: greet someone
      task: "echo hello {{inputs.name}}"
      tool: tool:bash
      depends_on: []

Run it with atelier run hello --input name=world.

A bigger conduit (deploy_pipeline)

A six-step pipeline that combines shell commands, an AI review, a human approval gate, retry loops, conditional branches, and a nested sub-conduit. It illustrates what is possible — a chatbot, a daily report, or an agent loop would look entirely different.

name: deploy_pipeline           # must match the folder name
description: Build test deploy
timeout: 3600                   # seconds per task, default 3600
max_concurrency: 3              # max tasks running in parallel, default 3

inputs:
  repo_url: The git repo URL
  branch: Branch to deploy
  env: Target environment

tasks:
  - clone_repo:
      description: Clone
      task: "git clone -b {{inputs.branch}} {{inputs.repo_url}} /tmp/build"
      tool: tool:bash
      depends_on: []

  - run_tests:
      description: Run tests
      task: "cd /tmp/build && make test"
      tool: tool:bash
      depends_on: [clone_repo]
      repeat: 3                          # try up to 3 times
      until: output.match(PASS)        # ...stopping early on success

  - code_review:
      description: AI review
      task: |
        Review /tmp/build/src for security issues.
        End your response with exactly one of:
        VERDICT: APPROVE
        VERDICT: REJECT
      tool: harness:claude-code
      depends_on: [clone_repo]
      interactive: false

  - approve:
      description: human gate
      task: "I need a final confirmation"
      tool: tool:hitl
      depends_on:
        - run_tests
        - code_review.output.match(VERDICT:\s*APPROVE)
      inputs:
        confirm: "Type 'yes' to approve deploy"
        reason: "Short reason for the decision"

  - deploy:
      description: Run deploy sub-conduit
      task: deploy_to_env
      tool: tool:conduit
      depends_on: [approve]
      inputs:
        target_env: "{{inputs.env}}"
        build_path: /tmp/build

  - rollback:
      description: Rollback if review rejected
      task: "make rollback"
      tool: tool:bash
      depends_on:
        - code_review.output.not_match(VERDICT:\s*APPROVE)

Step by step:

• clone_repo runs first because nothing depends on it.
• run_tests and code_review both wait on clone_repo, then run in parallel.
• run_tests retries up to 3 times, stopping as soon as the output contains PASS.
• code_review asks Claude Code to review the code and end with either VERDICT: APPROVE or VERDICT: REJECT.
• approve only runs if Claude approved (...match(VERDICT:\s*APPROVE)). It asks the human two typed questions on the terminal.
• deploy only runs after the human approves, and calls another conduit (deploy_to_env) as a nested run.
• rollback only runs if Claude rejected. The two branches are mutually exclusive — the unmet branch is silently skipped, not failed.

Conduit reference

A conduit has a name, a short description, an optional inputs map, and a tasks list. Each task has a name, a task body, a tool value, and a depends_on list.

Templating

• {{inputs.<name>}} — a conduit input or HITL answer.
• {{<task_name>.output}} — the printed output of an earlier task. The earlier task must appear in depends_on.
• {{loop.previous}} — this task's output from its previous loop iteration (empty before the first iteration completes). Only valid on a looping task (repeat > 1).
• {{loop.history}} — every prior iteration of this task, rendered as numbered blocks. Only valid on a looping task (repeat > 1).

A missing {{inputs.x}} fails the task immediately; a reference to a task that was skipped or hasn't completed skips the referencing task. atelier run rejects an --input or --input-file key the conduit neither declares nor references, so a mistyped key fails before the run starts.

Conditional dependencies

<task>.output.match(<regex>)        # dependency met if regex matches
<task>.output.not_match(<regex>)    # dependency met if regex does NOT match

The regex is everything between the leftmost ( and the last ). Python's re.search is used.

Quotes around the regex are optional and stripped when present, so output.match(PASS) and output.match("PASS") behave identically. To match a literal quote character, escape it — output.match(\"PASS\") looks for "PASS" with the quotes.

If a condition is not met, the task is skipped, not failed. Anything that depends on a skipped task is also skipped.

Loops (repeat + until / while)

A task with repeat > 1 can break out of its loop early:

until: output.match(<regex>)       # break as soon as an output matches
until: output.not_match(<regex>)   # break as soon as no output matches
while: output.match(<regex>)       # loop while an output matches; break otherwise
while: output.not_match(<regex>)   # loop while no output matches; break otherwise

Set at most one of until / while. The first iteration always runs before the predicate is checked.

For tool:conduit loops, the predicate sees every nested sub-task output of that iteration and fires on any match.

- retry_while_rate_limited:
    tool: tool:bash
    task: 'curl -s -o body -w "%{http_code}" https://api/x'
    repeat: 10
    while: output.match(^429$)

- run_until_test_passes:
    tool: tool:conduit
    task: build_and_test
    repeat: 5
    until: output.match(PASS)

Retries and per-task timeout

• retries: <n> — if a task fails, re-run it up to n more times (default 0). This is different from repeat, which loops a task that is succeeding.
• timeout: <seconds> — override the per-task time limit for one task. When omitted, the conduit-level timeout applies.

Asking a human (tool:hitl)

A tool:hitl task declares its own inputs: {name: description} map. At runtime flow-atelier prints the prompt, asks for each input by name on the terminal, and saves the answers so downstream tasks can use them as {{inputs.<name>}}.

Long AI conversations (interactive: true)

When a harness task sets interactive: true, flow-atelier appends this line to every message it sends to the AI:

When — and only when — you are completely finished, output the exact token [ATELIER_DONE] to signal completion.

Then it keeps the conversation open: the AI replies, flow-atelier streams the reply to your terminal, and if the AI didn't write [ATELIER_DONE] yet, flow-atelier asks you for the next message to send back. The loop ends when [ATELIER_DONE] shows up.

How that next message reaches flow-atelier depends on where the run started. On the terminal it reads one line from stdin — typed at the › cursor, or piped in for scripted runs. Under atelier serve the same conversation travels over /ws/run-conduit: the agent's prose is streamed to the client as it is written, and the client sends the reply back on the same socket.

Tool permission requests are automatically approved by default. Add a conduit-level interaction policy to choose human or supervisor decisions.

Non-interactive tasks run one turn and stop.

For a direct interactive agent conversation without writing a conduit, use atelier ask (Claude Code by default; --harness <name> picks another agent from atelier list harnesses):

atelier ask "Help me write a specification" --path /absolute/path/to/project

--path is required and becomes Claude's working directory. Claude's questions are read from stdin just like any other interactive harness task. The resulting flow is still recorded under the .atelier/flows/ directory from which you invoked atelier.

Conduit-wide human and supervisor policies

Choose the behavior once for every harness task in a conduit:

interaction:
  questions: hybrid       # human | supervisor | hybrid
  permissions: approve_all # approve_all | human | supervisor | hybrid
  supervisor:
    tool: harness:codex
    instructions: |
      Follow existing project conventions.
      Escalate scope changes and unspecified product preferences.

human always asks you. supervisor always delegates and fails if the supervisor cannot decide. hybrid lets it answer or escalate to you. approve_all automatically allows tool permissions. The supervisor receives the full worker session exposed through ACP, including prior replies and tool activity. Answers are attributed and recorded.

Adding interaction enables conversations for all harness tasks. Omitting it preserves existing task-level interactive behavior and automatic permissions. The designer exposes these controls in the conduit panel. Explicit tool:hitl gates remain human; nested conduits use their own policy.

See interaction policies for limits, failure handling, permission boundaries, and the complete configuration.

Interactive turns over the WebSocket

Three envelopes carry the conversation. They are additive — a client that ignores them still runs flows exactly as before.

From	type	Fields
server	agent_message	flow_id, task, text
server	agent_input_request	flow_id, task, request_id, prompt
client	agent_input_answer	flow_id, request_id, answer

agent_message is one chunk of agent prose, sent as the agent writes it, so the client can show the question before the request to answer it arrives. agent_input_request is the turn being handed back to you.

The request_id from a request must be echoed verbatim in the answer. That is what allows several interactive tasks in one flow to be waiting at the same time: each reply goes to the prompt that asked for it, not to whichever prompt happens to be pending. An answer carrying an unknown or already-used request_id comes back as an error envelope.

The socket is only the local transport between the UI and the atelier serve process on your machine. The harness itself still runs as that agent's own CLI under your existing login for it, so an interactive conversation over the WebSocket needs no LLM API key — exactly as on the terminal.

Where conduits live

Conduits can live in two places:

• Project: ./.atelier/conduits/ — scaffolded by atelier init.
• Global: ~/.atelier/conduits/ — shared across all projects.

When you run a conduit, flow-atelier checks the project folder first, then the global folder. A project-level conduit silently overrides a global one with the same name.

Flows are always project-local — every atelier run writes its flow folder under .atelier/flows/ in the current working directory.

Commands

# authoring
atelier init
atelier create <name> [--description <text>] [--template hello|code-review]
                                                       # scaffold a starter conduit
atelier check [<conduit>] [--json] [--recursive]        # validate conduit(s) without running
                                                       # --recursive also checks the conduits they call
atelier plan <conduit> [--json]                        # print the DAG as ordered waves, run nothing
atelier show <conduit> [--json]                        # print its definition and inputs, run nothing
atelier schema                                         # print the conduit.yaml JSON Schema for your editor

# <flow_id> below accepts a unique prefix, or 'latest' for the most recently started flow
# running
atelier run <conduit> [--input key=value ...] [--input-file key=path ...]
                      [--show-steps/--hide-steps]
                                                       # --input-file loads a UTF-8 text file into one input
atelier ask <query> --path <directory> [--harness <name>]   # interactive agent session (default: claude-code)
atelier run --resume <flow_id>                         # resume a failed/crashed flow
atelier run --again <flow_id>                          # fresh run reusing a past flow's inputs
atelier stop <flow_id>                                 # gracefully halt a running flow

# inspecting
atelier status <flow_id>
atelier wait <flow_id> [--timeout 60]                  # block until it finishes; exit 0 only if it did
atelier logs <flow_id> [--task <name>] [--follow] [--json]
atelier outputs <flow_id> [--task <name>] [--json]    # read back a finished flow's results
atelier timing <flow_id> [--json]                      # per-task duration, slowest first
atelier list conduits
atelier list flows [--conduit <name>]
atelier list schedules [--json]
atelier list harnesses [--ready] [--json]
atelier rm <flow_id> [--force] [--yes]                 # delete one flow run
atelier prune [--conduit <name>] [--older-than <days>] [--keep <n>]   # bulk-delete old flows

# sharing conduits (see "Installing conduit packages" below)
atelier install <source> [--ref <git-ref>] [--project] [--force]
atelier update <package>                               # re-fetch and re-install from source
atelier uninstall <package>                            # delete a package's conduits

# scheduling
atelier schedule add <file.{json,yaml}>
atelier schedule rm <id-or-name>
atelier schedule run-now <id-or-name>
atelier schedule history <id-or-name>
atelier schedule daemon [--reload-interval 30] [--log-level INFO]

# HTTP + WebSocket server
atelier serve [--host 127.0.0.1] [--port 8000] \
              [--reload-interval 30] [--cors-origin URL]* \
              [--log-level INFO]

# maintenance
atelier self-update                                    # prebuilt binary only; uv installs use `uv tool upgrade`

Installing conduit packages

A conduit is just a folder, so conduits are shareable. atelier install installs them from a git repo or a local path:

atelier install owner/repo                  # GitHub shorthand
atelier install https://github.com/owner/repo.git
atelier install ./some/local/package
atelier install owner/repo --ref v1.2.0     # pin a branch, tag, or commit

You are asked whether to install globally (~/.atelier) or into the current project (./.atelier); --project / --no-project answers that up front. An existing conduit of the same name is skipped, not overwritten, unless you pass --force.

Conduits are code. A conduit can run any shell command on your machine the moment you atelier run it. Read a package before you install it, and pin --ref for anything you don't control.

A package is any repo with its conduits under .atelier/conduits/ and an atelier-package.yaml at the root:

name: my-conduits        # letters, digits, _ and - only
version: 1
conduits:
  - deploy
  - nightly_report

Each listed name must be a directory under .atelier/conduits/. The whole directory is copied, so helper scripts and templates next to conduit.yaml travel with it. Without a manifest, flow-atelier discovers conduits by scanning that directory and warns that it did so. Schedules are never installed — they hold machine-specific state.

atelier update <package> re-fetches from the recorded source and re-installs. atelier uninstall <package> deletes only the conduits that install actually wrote, so a conduit that was skipped on collision is left alone.

Running on a schedule

atelier schedule daemon runs conduits on a wall-clock schedule. Each schedule is one YAML file under .atelier/schedules/<name>.yaml. The daemon is one foreground process you can put under systemd, launchd, or any supervisor.

To register a schedule, write a YAML file like the one below and run atelier schedule add <file>:

conduit_name: report
inputs:
  date: today
run_path: /abs/path
schedule:
  mode: recurring
  name: weekday mornings
  days: [1, 2, 3, 4, 5]
  times: ["06:00", "12:00"]

days are 1=Mon .. 7=Sun; times are "HH:mm" 24-hour strings. One-shots use mode: once with a run_at ISO datetime instead of days / times. Fixed intervals use mode: interval with every_minutes (e.g. every_minutes: 30 for every half hour, 120 for every two hours) — these repeat forever. atelier schedule add also accepts the same shape in JSON if you prefer that format. Like atelier run --input, it rejects an inputs key the conduit neither declares nor references, so a mistyped key fails at install time instead of silently running with the default on every fire.

• New or removed schedules are picked up on the next reload tick (default 30s).
• One-shot schedules remember they fired, so a daemon restart never re-runs them.
• Each schedule runs at most one instance at a time; missed fires are coalesced.
• atelier schedule run-now <id-or-name> fires a schedule immediately, bypassing the daemon.

HTTP API (atelier serve)

atelier serve boots a single process that hosts both the HTTP / WebSocket API and the scheduler daemon. It is the entry point the Flow Atelier visual frontend connects to.

Method	Path	Notes
GET	/conduits	List conduits
GET	/conduits/:name	Read one
POST	/conduits	Create (201 on success, 409 on collision)
PATCH	/conduits/:name	Partial update
DELETE	/conduits/:name	Delete
POST	/conduits/open-path	Reveal flow run path in OS file explorer
POST	/tasks/run	Run an ad-hoc one-task conduit
GET	/schedules	List active schedules
POST	/schedules	Create
DELETE	/schedules/:id	Soft-delete
GET	/flows	List prior flows
GET	/flows/:id/logs	Per-flow log entries
GET	/flows/:id	Tasks, dependencies and progress
GET	/flows/:id/tasks/:task/log	One task's rounds and actions, secrets masked
WS	/ws/run-conduit	Run flows, HITL + interactive AI turns
WS	/ws/flows/:id	The run page's feed: its map and one task's log, pushed as they change

Every run also has a page at /runs/<flow_id>: a map of its tasks with the ones running now framed, and the log of whichever task you click. It works for runs started from the CLI, the dashboard or the scheduler. The server follows the run's files and pushes each new line over /ws/flows/<flow_id> within a quarter second, shell output included.

atelier run and atelier ask print that page's address as soon as the flow starts, for example · run page http://127.0.0.1:8000/runs/20260925_0eb21391_hello, so an agent driving the CLI can hand the link to you. The page loads only while atelier serve runs from the same project directory as the run. Set ATELIER_SERVE_URL when you serve on another host or port.

Binds to 127.0.0.1:8000 by default; pass --host 0.0.0.0 to expose on the LAN — which requires ATELIER_API_TOKEN, see Security. --cors-origin is repeatable.

Conduits and flows resolve exactly as they do on the CLI — ./.atelier first, then ~/.atelier — so the conduits atelier init created in the directory you started the server from are the ones the UI shows. Schedules are the exception: they live in ~/.atelier/schedules/, since one daemon serves every project. Like atelier run --input, a run envelope on /ws/run-conduit whose inputs carry a key the conduit neither declares nor references comes back as flow_failed naming the key and its closest match, instead of starting the flow.

Security

The API runs shell commands on the machine hosting it, so treat reaching it as equivalent to a shell on that machine.

On loopback (the default). atelier serve binds 127.0.0.1:8000 and needs no token. Three guards keep a web page you happen to visit from driving it:

• Origin. CORS is restricted to localhost origins, never *.
• Host. Only localhost, 127.0.0.1, and ::1 are accepted as the Host header. This is what stops DNS rebinding, where an attacker's page resolves its own hostname to 127.0.0.1 so the browser treats the request as same-origin and sends no Origin for CORS to reject. Requests carrying any other Host get 400 Invalid host header.
• WebSocket origin. CORS does not apply to WebSockets, and a page on any site can open ws://127.0.0.1:8000/... with a loopback Host. So the WebSocket routes check the page's Origin themselves: they accept a local page, an origin passed with --cors-origin, or this server's own UI, and refuse any other with close code 1008 before the connection opens. A client that sends no Origin, such as a script, is not a browser page and is let through.

Anywhere else. Before binding to a non-loopback address, set ATELIER_API_TOKEN. Every REST request then needs Authorization: Bearer <token> and WebSocket connections need ?token=<token>; build the UI with a matching VITE_API_TOKEN so it can reach the authenticated API.

atelier serve refuses to start on a non-loopback host when ATELIER_API_TOKEN is unset. This used to be a warning that scrolled past in the same second the port opened, so an existing --host 0.0.0.0 setup with no token will now stop rather than serve:

$ atelier serve --host 0.0.0.0
error: refusing to serve on non-loopback host '0.0.0.0' without
ATELIER_API_TOKEN. Anyone who can reach this address could run shell
commands via the API. Set ATELIER_API_TOKEN, or bind 127.0.0.1 (the default).

Binding a specific host also adds that host to the accepted Host values; a wildcard bind (--host 0.0.0.0) cannot know which names reach it, so it accepts any Host and relies on the token — which is why the token is mandatory there rather than merely advised.

Tool arguments reach your terminal. atelier run prints the argument that identifies each tool call — the bash command, the file path, the search pattern — so the run is readable. Credential-shaped values (Bearer <token>, sk-/ghp_/xox-prefixed keys, --password/TOKEN= flags) are masked as *** on the way to the screen. This is a heuristic that reduces casual leakage, not a guarantee: it will miss a secret that does not look like one. The recorded logs under .atelier/flows/<id>/ keep the unredacted text, so treat that directory as sensitive and check what you are pasting before sharing a terminal transcript.

Conduits are code. See the warning under Installing conduit packages: running a conduit runs whatever shell commands it contains.

Folder layout

The .atelier directory lives in the working directory where atelier is invoked.

.atelier/
├── conduits/
│   └── <conduit_name>/conduit.yaml
├── schedules/
│   └── <schedule_name>.yaml                # one YAML file per schedule
├── scheduler_state.json                    # fired-once markers
└── flows/
    └── <flow_id>/                          # <YYYYMMDD>_<uuid8>_<conduit>
        ├── input.yaml                      # the inputs this run was given
        ├── logs.jsonl                      # append-only log, one JSON object per line
        ├── progress.json                   # live per-task status
        ├── outputs.yaml                    # per-task outputs (written as tasks finish)
        └── flows/
            └── <child_flow_id>/...         # nested tool:conduit runs

Contributing

For test instructions, the project layout, and internal architecture notes, see DEVELOPMENT.md.