mercury

Mercury

Self-hosted agent orchestration platform with channels, kanban state, audit, and TDD. Solo build.

Private repo

typescriptnode.jsreactsqlitemcpagentsaiclaude

Mercury kanban dashboard with project columns and active tasks — Kanban dashboard: single source of state

01 · Why

Agent work gets messy fast when prompts, channels, project state, permissions, and follow-up live in different places. The first iterations were scrappy: a Slack bot here, a cron there, no audit trail, no way to replay a session, every retry a fresh roll of the dice.

I built Mercury to make my own agent work manageable: kanban for state, audit logs for decisions, channel gateways for input, and worktree isolation for file changes. Solo project: design, code, runtime, and operations are all mine.

02 · What

A SQLite-backed kanban board is the visible state of the work. Cron heartbeats handle pickup and follow-up. Specialist agents use a Roman-pantheon naming scheme: Jupiter routes, while Mars, Vulcan, Minerva, Diana, Apollo, and Saturn cover React, .NET, review, research, fast build, and persistence.

Telegram is the conversation gateway, Slack is update-only, and the dashboard is the operator surface. Every meaningful operation leaves an audit row. Worktrees isolate file changes per task. Spend caps and approval gates were added after real failures, not as theory.

03 · Stack

TypeScriptNode.jsSQLiteReactopencode runtimeAnthropic SDKMCPTelegramSlackTDD

04 · Features

Eight specialist agents on a Roman-pantheon naming scheme, coordinated by Jupiter
Kanban-driven state across tasks, staging, review, archived research, and dashboard controls
Cron heartbeat loop for autonomous pickup, follow-up, and operational checks
Telegram conversation gateway, Slack update channel, and local CLI entry point
opencode runtime for broad model compatibility, plus Claude-specialist fallback paths
MCP tool surface for Claude paths and plugin tooling for non-Claude models
Broker-style spend caps, replay/audit thinking, and permission gates
TDD discipline across unit, integration, and smoke tests for runtime + gateway behaviour

05 · Gallery

Mercury audit view filtered to tool invocations from a successful agent run — Successful agent run: tool invocation trail

Mercury Hive cross-agent feed with high-signal agent updates — Hive: cross-agent signal feed

Mercury Skills tab with reusable agent instruction bundles — Skills: reusable agent instruction bundles

Mercury project kanban board with active maintenance tasks and priority filters — Project board: kanban task state

Mercury agent control centre showing recent agent sessions and outcomes — Control centre: session history

Mercury agent registry with enabled specialists, models, skills, and handles — Agent registry: specialists and models

Mercury queue controls showing concurrency limits and in-flight spawn capacity — Queue controls: concurrency limits

Mercury Hive mind feed of cross-agent task completion and escalation events — Hive: cross-agent operational feed

06 · Status

Running as a private, self-hosted control plane for my own projects. Maintenance-only now, because the product lessons are moving into Nido, but the orchestration, channel, audit, and reliability patterns have all been used for real work.