The control plane for autonomous AI coding

Ship issues autonomously.
Stay on budget.

AgentDock puts AI coding agents to work on your backlog — and keeps them accountable. Issues are scored before any build budget is spent, every run is capped, and every decision lands in an append-only audit ledger.

Get started See how it works

Self-hosted with Docker Compose
Your keys, encrypted at rest
Claude Agent SDK by default

acme/payments-api / Night wave Dispatching

Spend $38.40 of $60.00 cap

Window 22:00–07:00

Parallel builds 3 of 3

Pull requests 4 opened

IssuePreflightForecastStatus

#412Retry webhook delivery with backoff 92proceed $2.10 Merged

#418Add idempotency keys to refunds 88proceed $3.40 In review

#421Paginate the invoices endpoint 84proceed $1.80 Building

#427Migrate sessions to OAuth 2.1 58split — Deferred

#430“Make checkout faster” 31clarify — Held back

02:14:07 pr_opened #418 · cost $3.12, under its $3.40 forecast

Works with the stack you already run

GitHub
Azure DevOps
Jira Cloud
Claude Agent SDK
Cursor
MCP clients

Why AgentDock

Agents are easy to start. Hard to trust.

Autonomous agents fail in three expensive ways: they start on work that was never clear, they keep spending after they should have stopped, and they leave no record of why. AgentDock closes all three — so you can hand over the backlog, not just a prompt.

Ship while you sleep

Queue the backlog at the end of the day, set a budget and a window, and step away. The dispatcher works through it with bounded concurrency — you review pull requests over morning coffee. No babysitting, no overnight pager.

18:40Wave queued: 9 issues, $60 cap
22:00Dispatch starts, 3 in parallel
02:14PR #418 opened, checks green
07:006 PRs ready · $41.20 spent

Hard budget caps

A ceiling is a ceiling. Runs that would breach it never start, and runs that reach it stop at a clean boundary. Your AI spend stays as predictable as any other line item.

Cut wasted runs

Preflight reads every issue before a line of code is written and decides whether it is ready. Pay cents to assess a bad issue instead of dollars to fail at it.

proceed
clarify
split
defer

Multi-provider

GitHub and Azure DevOps side by side, Jira Cloud as an issue source, and the agent backend you choose per repository — under one budget model and one audit trail. Adopt AgentDock without re-platforming or locking yourself in.

Sources

GitHub
Azure DevOps
Jira Cloud

Engines

Claude Agent SDK
Cursor Cloud
Anthropic-compatible

How it works

From backlog to reviewed pull requests in three moves.

No new workflow to learn. AgentDock works on the repositories and issues your team already has.

Connect your repo

Import a GitHub or Azure DevOps repository. AgentDock syncs its issues, warms up a readiness score and stores every credential encrypted at rest with pgcrypto.

GitHub · Azure DevOps · Jira Cloud
Set a build wave

Choose a spend ceiling, a time window and how many builds run in parallel. Every issue gets a cost forecast up front, and nothing starts that would break the cap.

Budget · window · 1–16 parallel builds
Sleep. Wake to PRs.

The dispatcher preflights, ranks and builds in order, then opens pull requests that carry a reviewer guide and the evidence behind every claim. Every transition lands in the audit ledger.

Preflight → rank → build → pull request

Guardrails

Every dollar passes a gate.

Spend control is not a report you read in the morning. It is a stack of gates every run has to clear — before it starts, and for as long as it runs.

Before it starts. The cost forecast is checked against budget policies, restricted labels, risk tiers, the wave budget, the project’s lifetime budget and the global cap — the last two committed atomically with the run, so concurrent spawns can’t overshoot.
While it runs. Per-run cost and runtime ceilings, a loop cap and a stuck-loop detector stop a runaway session at a clean boundary.
Always. One kill switch stops new runs from starting, and a scheduled reconciler settles every budget reservation.

Spawn check #421 Paginate the invoices endpoint forecast $1.80

Kill switchruns_paused = falsepass
Global spend cap$412 of $2,000pass
Budget policyno restricted labelspass
Project lifetime budget$212 of $1,500pass
Wave budget$38.40 of $60.00pass
Run ceiling$1.12 of $8.00 · 14 minwatching

budget_reservation_reconciled every reservation settles to the cent

Reviewable by design

Pull requests that arrive with receipts.

An agent saying “tests pass” is not evidence. AgentDock pull requests carry a reviewer guide — what the agent assumed, what it could not verify, where to start reviewing — and flag any claim no command in the run backs.

Reviewer guide and “Not verified” flags in the PR description, and a run page that links every check to its evidence.
Second-model review surfaces P0 findings before a human looks.
A PR babysitter follows CI, review and merge state, and queues approval-gated fixes when something breaks.

Retry webhook delivery with exponential backoff #418 Openagentdock/412-webhook-retry → main · $3.12

Reviewer guide

Assumed: Receivers treat retried deliveries as idempotent.
Not verified: Behaviour against the staging webhook relay.
Start here: src/webhooks/deliver.mjs — retry schedule and jitter.

Verification

Tests48 passed
Lint & typecheckclean
Second-model reviewno P0 findings
Secret scanclean
PR checks2 of 3 green

Platform

Everything a control plane for agents needs.

One dashboard, one REST API and one background worker over a single Postgres database.

Preflight analysis

Scores every issue on ambiguity, scope risk, testability, conflict and security surface, and forecasts its cost before it reaches the build queue.

Build waves & autonomous dispatch

Time-boxed waves with their own budget, a curated queue you can dry-run, goal-driven sequencing and up to 16 parallel builds.

Layered cost guardrails

Run, wave, project and global ceilings plus a kill switch. A run that breaches its limit cancels itself.

Smart issue ranking

A multi-term score with configurable strategies and per-repository weights decides what ships first.

Second-model review

A second model reviews the diff and posts any P0 findings on the pull request — optionally holding it as a draft.

PR babysitter

Tracks CI, reviews and merge state after the pull request opens, with approval-gated fixes and opt-in, geofenced auto-merge.

Learning loop & memory

Per-repository memory and playbooks shape each prompt, and every outcome feeds back into calibration.

Append-only audit ledger

Every invocation, dispatch decision and skip reason is recorded with actor, timestamp and outcome.

Encrypted credential storage

API keys and tokens live in Postgres, encrypted with pgcrypto. No plaintext secrets in .env.

ROI analytics

Cost per merged PR, time to merge, waste patterns and forecast accuracy — measured from your own history.

MCP server

Claude Code, Cursor or any MCP client can query runs and operate build waves, under the token owner’s own permissions.

Roles & API tokens

Four system roles plus custom roles built from granular permissions, and revocable tokens for machine access.

Self-hosted

Your infrastructure. Your keys. Your ledger.

AgentDock runs next to your code as a Docker Compose stack over one Postgres database. Your credentials and your audit trail live in that database, under your control.

Encrypted at restEvery API key and PAT, with pgcrypto.
Role-based accessSystem and custom roles, per permission.
Strict by defaultCSP, frame denial and HSTS behind TLS.
One kill switchStop new runs from Settings.

Quick start

# 1. An encryption key for APP_SECRETS_KEY in .env
$ openssl rand -hex 32

# 2. Database, schema and services
$ docker compose up -d postgres
$ docker compose run --rm agentdock-cli npm run db:setup
$ docker compose up -d dashboard-api nginx agent-worker

Dashboard on localhost:3000

Put your backlog on the night shift.

Connect a repository, set a budget and let AgentDock work through it — every run scored, capped and on the record.

Get started Sign in

Ship issues autonomously.Stay on budget.