Skip to content
Why Detent

You are not steering an agent. You are running your process at scale.

This page is the argument, the two default choices that define Detent's footprint, and the comparison matrix — including the rows where Detent loses.

01 · The inversion

A system, not an agent.

Autonomy-first tools put the intelligence in the agent and ask you to supervise it. Detent puts the intelligence in your spec and supplies the discipline. Same task, two interaction models.

“add OAuth token rotation” — autonomy-first assistant
  1. It starts as a prompt. The agent picks a plan.
  2. You read the plan and inspect partial edits.
  3. It misses the migration. You redirect.
  4. It skips the tests. You redirect again.
  5. You are the supervision loop.
“add OAuth token rotation” — Detent
  1. It starts as an issue naming the storage change, CLI behavior, migration, rollback, and tests.
  2. A worker executes that contract in an isolated worktree.
  3. It produces a reviewable pull request.
  4. It does not merge until your gates are green.
  5. Your process runs. You read one PR.

The goal is not to replace engineers or hide work behind opaque behavior. It is to scale the judgment of engineers who already have a high bar. The system does not try to be smarter than you; it tries to be as disciplined as you would be, every time, in parallel.

02 · Who this is for

And who should close the tab.

The engineer who runs delivery

Staff, principal, or lead. You already have CI gates, review standards, merge rules, and dependency ordering written down. Your bottleneck is throughput on well-specified work without lowering the bar.

You want the state machine, the gate model, the merge-train guarantee, and the isolation story. You will read a detent.yaml before you read a testimonial. There is no testimonial on this site.

The solo builder with five repos

One person, several repositories, a queue of work that is clear but unwritten. You want parallelism without babysitting five chat sessions.

Fast time to first run, a single-binary install, multi-project from one host, and visible budget and rate-limit state. Detent is free and MIT, and it bills agent work against a ChatGPT plan you already pay for.

The platform or devprod team

Evaluating agentic delivery for an organization. Your questions are governance questions.

Self-hosted, air-gappable, no vendor control plane, budget caps, fleet governance, and an audit trail in GitHub where reviews already live. MIT licensed.

Not for

People who want an autonomous assistant to talk to. People who want a hosted SaaS with a signup button. People looking for a coding copilot inside an editor. Detent is none of those, and it is better for everyone if that is clear on the home page rather than three pages into the docs.

03 · The defaults

Two choices define the footprint. Both are deliberate.

Why GitHub Projects, not Linear

The reference design Detent grew from polls a Linear board while code, pull requests, and CI live in GitHub — two systems for one unit of work. That split forces you to map Linear issue IDs onto GitHub PR numbers and to read discussion in two places.

Detent puts the whole state machine in one system. A GitHub Project is the board; its issues are the work items, its pull requests are the deliverables, and its comments and reviews are where every conversation happens. One ID space, one place to look.

Cost
GitHub Projects has no per-seat charge. Linear's Business plan is $16 per user per month — about $9,600 a year at 50 seats — and its free tier is capped.
API headroom
Authenticated GitHub REST allows 5,000 requests an hour, more for GitHub Apps. Linear allows 1,500 an hour against a complexity budget. A poller driving many repositories wants the larger ceiling.

Why Codex, not Claude

Detent dispatches agents non-interactively, headless, many at once. The important question is how that mode is metered.

  • A ChatGPT plan — Plus, Pro, or Business — covers Codex CLI usage including scripted codex exec automation, billed against the subscription you already have.
  • Claude Code keeps interactive terminal and IDE use on subscription limits. Effective June 15, 2026, Anthropic moves headless claude -p, the Agent SDK, Claude Code GitHub Actions, and third-party Agent SDK apps to a separate monthly Agent SDK credit — per-user, non-rolling, with overages at standard API rates.

For an orchestrator running agents around the clock in parallel, the Codex subscription is the default that makes the economics work. Detent still supports explicit backend routing, including Claude Code, per role — so operators can pick a backend where the limits, auth mode, and isolation trade-offs fit that work.

04 · Comparison

Including the rows Detent loses.

This matrix is copied from docs/comparison.md in the repository. Where a competitor is better, the table says so.

Capability comparison between Detent and adjacent agent orchestration tools.
CapabilityDetentOpenAI SymphonyCopilot agentCursorHermesOpenClawHyperagent
Self-hosted, no vendor control plane
yes
yes
no
no
yes
yes
no hosted
Runs fully local / air-gappable
yes
yes
no
no
yes
yes
no cloud
Board/tracker-native (issue→PR)
yes GH Projects, issue fields, or labels
yes Linear
yes GH Issues
no
no
no
no own workspace
Deterministic gated merge train
yes
partial per spec
no
no
no
no
no
Budget / cost caps
yes
no
partial
partial
no
no
yes hosted controls
Multi-project
yes
no
yes
yes
no
no
partial workspace-scoped
Multi-instance fleet governance
yes
no
no
no
no
no
partial hosted agent controls
Model-agnostic, BYO incl. local
partial codex now, seam shipped
no Codex
yes vendor-managed
yes vendor-managed
yes
yes
no cloud-managed
Local skills / workflows (your e2e etc.)
yes
partial
no
partial
yes
yes
yes hosted skills/knowledge
Multi-channel triggers
partial tracker-driven
partial Linear
partial GitHub
partial IDE/cloud tasks
yes messaging gateway
yes local gateway
yes Slack, schedules, webhooks, email, Telegram, Live Mode
Open source
yes MIT
yes Apache-2.0
no
no
yes MIT
yes MIT
no closed-source
Free (bring your own model cost)
yes
yes
no paid
no paid
yes
yes
no usage-billed
Single static binary
yes
no Elixir/BEAM
not applicable SaaS
not applicable SaaS
no gateway
no gateway
not applicable SaaS
~5-min setup
yes
no
yes zero-install
yes
partial
partial
partial hosted onboarding

✓ yes · ◐ partial · ✗ no · — not applicable. Last verified July 6, 2026; check vendor pricing and models before relying on them.

Detent ↗

Single-binary Go orchestrator for GitHub-native issue-to-PR work, using ProjectV2 or boardless status sources.

OpenAI Symphony ↗

Detent's origin point: an Apache-2.0 spec plus an Elixir reference implementation for Codex on Linear.

Copilot agent ↗

GitHub's paid issue- or prompt-to-branch-and-PR cloud agent.

Cursor ↗

IDE-first agent product with cloud and background agents, automations, and optional self-hosted workers.

Hermes ↗

Nous's MIT personal assistant with memory, skills, model providers, and a messaging gateway.

OpenClaw ↗

MIT personal assistant centered on a local gateway and cross-channel automation.

Hyperagent ↗

Airtable's closed-source, cloud-hosted agent platform for persistent agents with identity, tools, and knowledge.

Copilot and Cursor have closed much of the “runs near my code” gap and win zero-install inside their own platforms. They do not give you a board-native release runtime under your control. That is the trade, stated plainly.

05 · Objections

The questions people actually ask.

Will it merge garbage?

Not without passing the gates you wrote. Promotion to Merging requires whatever the workflow declared — a human release, a current-head automated review, or a linked PR with green CI and a quiet window. Then the serialized train rebases, watches CI on the exact commit about to land, and merges one candidate at a time.

What if the agent goes off the rails?

It works on its own branch in its own worktree, and it cannot merge — only the train merges, after your gates clear. Be precise about the limit, though: a worktree is collision isolation, not an OS sandbox. Detent's own config runs the agent with full host access, so the guarantee is about what lands, not about what the process can touch.

Why Codex and not Claude?

Metering, for headless parallel dispatch. See the defaults above. Detent supports explicit backend routing including Claude Code, per role.

Why GitHub Projects and not Linear?

One ID space and one place to look, plus no per-seat cost and 5,000 requests an hour against Linear's 1,500 on a complexity budget.

Is this a toy?

Judge it the way you would judge any repository: version history, a CI badge, a release channel, installer smoke tests on Ubuntu and Windows against real release assets, and the fact that it builds itself.

What is the catch?

You bring the model cost, and you write the contracts. Detent will not invent your acceptance criteria, and it is not trying to. A vague issue produces a vague pull request faster than before.

Keep reading