CLI · MCP server · Claude Code hooks
“The agent said it’s done” is never the evidence.
ACLine is a local SDLC tracker for coding agents. Specs, tasks, decisions and an append-only audit trail live in one SQLite file — a task only completes when the store holds the checks and approvals its risk requires.
curl -fsSL https://acline.nizaami.com/install.sh | bash
What the store enforces
Same rules for a person and an agent.
One store
A single SQLite file (~/.acline/store.db), shared by every project you track. No servers, no sync.
A completion gate
task done refuses without a recorded, passing check — and for high risk, without a person's approval.
A tamper-evident trail
Events are hash-chained; approvals and checks are sealed. acline verify says what was altered.
Agents propose, not decide
Agents draft specs, plans and checks; a person — or the holder of an approval token — approves and accepts.
A guard for Claude Code
A PreToolUse hook that fails closed and refuses secret files, destructive commands and writes outside the project.
MCP server included
acline mcp serve exposes the store to any MCP client, including the VS Code extension.
Quickstart
acline init registers the project and scaffolds hooks, skills and a vault.
acline dashboard shows what needs attention;
acline next says what should happen to a task and who should do it.
- Platforms
- macOS and Linux. On Windows, use WSL — the Claude Code hooks need a POSIX shell.
- Requirements
- Just the
aclinebinary onPATH— no Python dependency.
# register the project, scaffold hooks + CLAUDE.md
acline init
# enable the approval token (run this yourself)
acline auth init
# store, audit trail, token, guard hooks, claude binary
acline doctor
# an agent may draft; a person approves
acline spec add "Export as CSV" --body "..."
acline spec approve 1
acline task add "CSV writer" --spec 1 --risk medium
# runs the tool and records its real result
acline check run 1 --kind test
# the gate decides
acline task done 1
Trust in one paragraph
Until you run acline auth init, identity is whatever
ACLINE_ACTOR_TYPE says. Enabling the token closes that: approvals, gate
overrides, accepting or retiring decisions, approving specs and plans, loosening a task's risk or autonomy, and
importing a snapshot all then need a secret the agent is never given. The guard is a best-effort denylist, not
a sandbox — see the
threat model
for the full list of limits.