Agents write the code. An engine runs the DAG.

CodeDAG is the orchestration engine for coding agents. Your prompts, linked by dependency, are the DAG. However many tasks, it all runs exactly as written: no babysitting, no model keeping track.

Join the waitlist Read the docs

Runs locally on your existing agent subscriptions · Free credits during the private beta

.codedag/dags/site-search.yaml

$ ✔ Submitted successfully: .codedag/dags/site-search.yaml

Without CodeDAG

Until now, that engine has been you. Step away, and it all stops.

Several terminals, each session waiting on your next prompt. You track what depends on what, start each task by hand, and merge the branches yourself.

With CodeDAG

01
Orchestrate
Deterministic scheduling. Autonomous agents.
Dependencies are declared, and the engine schedules them in topological order: upstream finishes, downstream starts. What runs next is the DAG’s call, never a model’s. Each task is just a prompt: a goal is enough, no steps required.
- Fresh session, fresh context per task
- Orchestration as code: validate, rerun
- Or have an agent write the DAG
.codedag/dags/site-search.yaml
tasks: - id: 1 title: Write the spec agent: claude prompt: |Claude Code Read docs/search.md, write docs/spec.md. - id: 2 title: Write the tests dependencies: [1]engine agent: codex prompt: |Codex Write tests from docs/spec.md. - id: 3 title: Build the API
CodeDAG engineWorkspace 1Workspace 2Workspace 3Write the specBuild the APIBuild the UIWrite the testsMake tests passReview the diff
Yousubmitawayback

Steer

See every step. Stop any task.

Every task is a session you can attach to: watch it live, correct it directly. Pause it, kill it, recover it. It picks up in the same session, context intact.

Full logs, tailed live
Stop one task or the whole workflow
Review gates where you want them

cdg dag status site-search --watch

Submitted Tasks (5):
  🔵 [done]        [done]  Task 2: Write the testsNeeds your review  🔴 [stopped]     [killed]  Task 3: Build the APISession kept  🟢 [in-progress] [working]  Task 4: Build the UI  🔘 [pending]     [blocked]  Task 5: Make tests pass  🔘 [pending]     [blocked]  Task 6: Review the diff
Reviewed Tasks (1):
  🤖☑️ [auto-accepted]  Task 1: Write the spec
Progress: 2/6 tasks completed (33%), 1 reviewed

cdg task view 4 -w site-searchClaude Code

›use the existing Table

Pause
Kill
Recover

$ cdg task recover 3 -w site-search

03
Merge
Auto-merge. Auto-resolve. main stays untouched.
One branch per task. Upstream results merge into downstream tasks automatically, and agents resolve the conflicts. Drop or redo any single task. It all ends on one integration branch; landing that on main is your call.
- Parallel tasks never share a checkout
- Every task’s diff reviewable on its own
- Dropped work is kept as a snapshot

The eight principles behind these choices

What it’s for

If you can draw the DAG, you can hand it off.

Spec, build, verify
Settle the spec, build the parts in parallel, then make it all pass.
Several changes at once
Independent changes each get their own branch and merge at the end.
One builds, one reviews
Claude Code writes it, Codex reviews and fixes it. Or the other way around.
Step after step, overnight
A big refactor as a dozen sequential tasks, each building on the last. No babysitting.

Compare with Claude Code Dynamic WorkflowWorks with Claude Code and Codex today.

Data privacy

Not a line of your code reaches us.

Your machineYour repoagents read and write hereClaude CodeCodexcdgthe CodeDAG CLIstarts sessionscode and contextsign-in and credits

Never sent to CodeDAG

your code
run logs
session contents

Model providersDirect, as always. Your own accounts.

CodeDAG account service

Only what sign-in and credits need

The workflow file is discarded once counted, never stored.

Your machineYour repoagents read and write hereClaude CodeCodexcdgthe CodeDAG CLIstarts sessions

Never sent to CodeDAG

your code
run logs
session contents

code and context

Model providersDirect, as always. Your own accounts.

sign-in and credits

CodeDAG account service

Only what sign-in and credits need

The workflow file is discarded once counted, never stored.

Never uploads your code
Your code, logs, and sessions are never sent to CodeDAG. It only gets your sign-in, the workflow file, and basics like the CLI version.
Never replaces your agents
It runs your own agents, on the subscriptions you already have. No token bill.
Never touches your main branch
It doesn’t commit to main, and it doesn’t push to or pull from your remote.

Pricing

Billed by the task. One credit each.

This workflow has 6 tasks, so a run costs 6 credits. However long it runs, however many rounds of rework, it’s still 6.

See pricing and common questions

$12/month: Or $96 a year: $8/month, 33% off
800credits/month: About 800 commits, over 25 a day
800free credits: When you sign up. Valid for 30 days, no card required

Model usage stays on the agent subscriptions you already have. CodeDAG never bills by the token.

The private beta opens soon. Save your spot.

Invites go out by email, in batches. Everyone invited gets free credits while the beta runs.

Join the waitlist

Read the docs first

Agents write the code. An engine runs the DAG.

Until now, that engine has been you. Step away, and it all stops.

Deterministic scheduling. Autonomous agents.

See every step. Stop any task.

Auto-merge. Auto-resolve. main stays untouched.

If you can draw the DAG, you can hand it off.

Spec, build, verify

Several changes at once

One builds, one reviews

Step after step, overnight

Not a line of your code reaches us.

Never uploads your code

Never replaces your agents

Never touches your main branch

Billed by the task. One credit each.

The private beta opens soon. Save your spot.

Auto-merge. Auto-resolve. `main` stays untouched.