What it is in one line
An open-source, self-hostable platform that turns coding agents into visible, reviewable,
recoverable delivery workflows: agents run in real Docker sandboxes, swap backends per need
(Cursor / Claude Code / CodeBuddy / Trae), pause for human approval at critical nodes,
exchange structured artifacts, and finally open the PR/MR themselves.
Who built it
Solo-led by cocofhu. MIT license, repo created 2026-07-24, 77 stars / 3 forks. Backend in Go
(FSM engine, API, SQLite, artifact MCP, scheduling, audit), frontend in Vue 3 + Vue Flow (the
orchestration canvas), sandboxing via Docker plus a custom sandbox-gateway. ~723 commits,
currently at 0.2.1-beta. The commit history includes AgentBot and Cursor commits — the author
is literally building this with agents.
Read: a tool whose job is "make agent delivery trustworthy" is being built by agents itself.
That self-certification chain is informal but oddly consistent. 77 stars says people who looked
nodded; it has not been validated yet.
What it actually does
- Visual FSM orchestration → draw agent/human-gate/rollback nodes on a canvas; edges
define success, failure, and rollback paths
- Human-in-the-loop gates → pause at critical nodes so a human can approve, reject, or
request changes from structured artifacts
- Real Docker sandboxes → isolated execution via sandbox-gateway; each agent picks its own
backend, so one workflow can mix different agents
- Artifact contracts + MCP → every run gets its own artifact MCP and token; agents exchange
structured outputs via write_artifact / set_* / node_complete, isolated per run
- Git delivery → GitHub/GitLab/SSH credentials injected into the sandbox; auto-opens
PRs/MRs
- Self-hostable → one-repo Docker Compose bring-up, images published on GHCR, no local
build needed
- Typical workflow → Clarify → Research → Proposal → human approval → Plan → Implement →
Test → Review → human confirmation → PR/MR
What old behavior it replaces
Handing delivery to a coding agent used to default to post-hoc review: the agent finishes, opens
a PR, and a human eyeballs it — while what the agent changed and why was often a black box, so
the reviewer had to re-read the whole diff to judge it.
Approving replaces that chain in two shapes. First, it moves review earlier: approval happens
before the code exists — the human approves the Proposal and the plan, not the aftermath.
Second, it makes rollback an explicit path in the graph instead of a human detective job.
It also replaces a more basic behavior: agent tools were siloed (Claude Code's way, Codex's
way), with no platform layer to orchestrate different agents into one gated flow. Approving
wants to be that orchestration layer.
Read: "approve before code is written" is the right design. Reviewing a finished PR is the
most expensive form of review — the work is done, overturning means redoing. Approving a plan
is a far cheaper checkpoint.
Business model
Not disclosed. MIT open source, self-hostable, no cloud service, no pricing page. The
homepage shows team-shaped features (parallel-sprints, a PM that chats) but nothing is priced.
Read: the same classic script as its peers — open source for distribution, a future cloud for
revenue. But cloud, enterprise tier, and pricing are all absent, so the model has not started.
Hard numbers
- 77 stars / 3 forks. Repo created 2026-07-24, ~723 commits, at 0.2.1-beta
- Four agent backends supported: Cursor, Claude Code, CodeBuddy, Trae
- Go backend + Vue 3 frontend + custom sandbox-gateway; Docker Compose self-host with
multi-GB sandbox images on first pull
- Users and revenue: not disclosed
Four-way read
| Dimension |
Call |
| Founder-product fit |
Background unknown, but building a "manage agents" platform with agents is unusually self-consistent tool-scenario fit |
| Product insight |
Two real problems: approval should move to the plan stage, and rollback should be an explicit path, not post-hoc forensics |
| Execution quality |
Complete architecture (Go backend + canvas frontend + sandbox gateway + artifact MCP + audit), 0.2.1-beta shows fast iteration |
| Timing |
Right. Coding agents already run; "how do I trust what they deliver" is exactly the next gate for enterprise adoption |
The call
This is the trust layer for agent delivery, positioned at the right moment, but not yet
validated.
Once single coding agents work, the enterprise question becomes "how do I trust what it
changed." Approving answers part of it: make agent workflows visible, approvable, and
recoverable — and move the human checkpoint from "review after it's done" to "review the plan
before it starts."
The transferable rule: where you place the approval point in a process decides its cost.
Reviewing a plan is an order of magnitude cheaper than reviewing a result — a result is
already committed, overturning means starting over; a plan is text, changing it takes minutes.
Deciding "at which step does a human say yes" is the highest-leverage design decision in any
automation.
Its weakness is the orchestration layer itself. Four backends, each with its own
acpBackend; the platform manages sandboxes and credentials. Adopting it requires willingness
to self-host, pull multi-GB images, and maintain your own sandbox infrastructure. That makes
it a tool for technical teams, not a sales-driven product.
Where it sits vs peers: AgentGate and Lelu build the permission/policy layer (approving
agent actions); Approving builds the delivery-flow layer (turning agents into gated
pipelines). The former is closer to a security tool, the latter to CI/CD for agents. Both
directions are valid; Approving walks the dev-delivery lane.
What to watch next
① Whether stars pass 300 in three months and any company publicly says it runs this — the line
between "watched" and "trusted"
② Whether a hosted tier or team pricing appears — the first step from free self-host to a
business model
③ Whether the backend list keeps growing as new agent tools ship — an orchestration layer
that lags new agents is dead
What you can take from it
Product logic: when designing human-machine workflows, decide where the human approval
point sits before deciding features. Copy this ordering: approve the plan (cheap) → execute →
review the result (expensive). Back plan approval with structured artifacts so the approver
has something concrete to read, and back result review with explicit rollback paths.
Pricing structure: none. Not disclosed.
Verdict
Worth watching, but early. Right position, right design (approval moved forward), fast
iteration — this is the closest thing in this batch to a shippable form of the "agent
governance" thesis. But 77 stars and zero business model mean it hasn't crossed from tool to
product. Revisit in three months against the checks above.