AcahtiDocs
EN中文GitHubSelf-host
Menu

DocsGet started

Overview

What Acahti is, how an agent delivery loop runs on it, and who does what.

Acahti is a self-hosted delivery control plane for coding agents. One instance gives your team Git hosting, pull requests, CI with structured check results, and npm and PyPI registries behind a single identity and a single MCP endpoint. Agents run the delivery loop; people set policy and watch.

Product and instances#

  • The product lives here, at acahti.benchyard.com: documentation, the agent skill, the install guide.
  • An instance is one deployment on your own infrastructure, such as https://acahti.example.com. It serves the web app, MCP at /mcp, the REST API at /acahti/v1, Git over HTTPS, and package registries.

Agents install the skill from the product site once and connect to whichever instance they work on. Everything instance-specific (URL, org, Git identity, admin and creation rights, installed plugins) comes from the instance's whoami.

The loop#

repo_get → branch_list → edit and push
                           ↓
                     checks_wait
                      ↙       ↘
           failed-step logs    green checks
                ↓                   ↓
            fix / rerun       review → pr_merge

The agent decides and repairs. Acahti supplies authenticated operations and results. pr_merge succeeds only when the newest pipeline on the PR head is green. A green pipeline means the configured checks passed, not that the software has no defects.

Who does what#

Agent Person
Connects with a personal Access Token (MCP) or an access token (CLI) Invites members, sets branch protection
Pushes, waits on checks, reads failed logs, reruns Owns pipeline secrets and repository access
Opens and merges pull requests, publishes packages Watches the Board and Inbox
Reports blocked runs Approves gated deploys (deploy_approve)

Aidy answers with cited sources and carries out actions you confirm, inside Acahti, using an administrator-configured model. External coding agents connect through the same authenticated MCP catalog. Bring Claude Code, Codex, Cursor, or any MCP client.

The web app#

The header carries the organization and a Dashboard | Agent switch on the left, the sections in the middle (Board, Repos, Pipelines, Packages, Tickets, Projects, Knowledge, Argos, installed plugins and Track), and search, the sidecar, notifications and your account menu on the right. The account menu holds your display name and email, access tokens, Admin (for administrators), language, theme, this documentation and sign-out. Collapse the sidebar to a narrow rail of monograms with its control or Ctrl/Cmd+B. Repos lists each repository with its language, stars, forks, open pull requests, last commit, latest pipeline and topics, filtered by team, type and language and sorted by last update, name or stars; organization administrators archive a repository to make it read-only and hide it from members (repo_archive). The sidecar holds Aidy and your chats; Agent mode makes it the main surface. See Aidy and global search and Track.

Under the hood#

An instance is a Go gateway in front of two mature kernels: Forgejo for Git, pull requests, and packages, and Woodpecker for CI. The kernels stay private; the gateway is the only public surface. Pipelines run on a separate Runner host. See How it works.

Next#

Plan, build, verify and learn#

Planning is shared by people and agents. The system measures (scope against the baseline, overdue and deferred work, capacity, burn-up) in one read, planning_brief; Aidy turns it into a status update, proposes how to close an iteration, drafts a delivery's changelog and fills in the day's time from your activity, and every change it proposes waits for your confirmation.

  • Projects: people, agents, repositories and work in one ordered list, each with an overview of risks, load and recent deliveries.
  • Product iterations: a scope baseline when work starts, out-of-plan changes, capacity and burn-up, a close-out that accounts for every unfinished todo, and deliveries that name their todos.
  • Track: follow a ticket, todo, iteration or delivery from report to release.
  • Knowledge base: shared documents and uploaded files, structured folders, rich editing, search and revision recovery.
  • Aidy and global search: search from the header, ask Aidy in the sidecar and follow citations to original records.
  • Argos: source-visible cases, shared plans, execution evidence and agent access.
  • Work and notifications: tickets, todos, time entries and owner notifications.

For agents: overview.mdEdit this page