SharperFlow Open Source

Vision

One Go daemon runs every MCP server: process supervision, hot reload, health checks, and an admin API your agent can call.

View on GitHub Read docs

quick install

curl -fsSL https://raw.githubusercontent.com/Sharper-Flow/Vision-MCP-Manager/trunk/scripts/install.sh | bash
vision daemon start -d

Single binary. A catalog of 15 pre-configured servers ships with it — context7, fetch, filesystem, github, sqlite, and more.

the problem

MCP servers, unsupervised

Every project wires its own MCP configuration. Servers crash silently and stay down. Nobody can see which tools are live, and the only fix is another per-project config edit.

Scattered configuration, duplicated across projects and machines
Silent crashes with no restart and no visibility
Static tooling — agents cannot adapt without a human editing config
stdio servers that cannot be shared across clients

proof

Real output from a running daemon

Captured from a live daemon. The per-server view is the Admin MCP surface, not the CLI.

vision daemon status

Daemon is running (PID 1774)
  Status: ok
  Uptime: 10h46m1s

vision config validate

Configuration is valid

vision_status — admin mcp

{"healthy": false, "uptime": "10h46m18s",
 "servers": {"running": 10, "stopped": 2, "error": 0},
 "memory_mb": 132.7}

Two servers are stopped, so the fleet verdict is false even while the daemon process is healthy.

vision_list — admin mcp

context7  running  :6276  uptime 11h9m0s
episode   stopped  reason: process stopped
lgrep     running  :6278  uptime 11h9m0s
svelte    running  :6279  uptime 11h9m0s

how it works

One daemon, four jobs

Centralized registry

One YAML file, ~/.config/vision/servers.yaml, defines every server. One place to look, one place to change.

Process supervision

Erlang-style supervision restarts crashed servers with exponential backoff from a 1-second initial delay.

Hot reload

vision daemon reload applies configuration changes, and your coding session keeps running.

Agent self-management

Agents add, remove, and restart servers through the Admin MCP API, within the catalog you pre-approve.

under the hood

Built for long sessions

Vision multiplexes each server across sessions and gives each session its own MCP process on a dedicated port. If a subprocess is reaped or crashes, the next tool call transparently spawns a fresh one.

The daemon probes reachability and derives an effective status per server, so a process that is up but not answering is not reported healthy.

Catalog: 15 pre-configured servers, from context7 to sqlite and puppeteer.
Transports: stdio servers locally, HTTP proxying for remote servers.
Surface: Go 1.24 single binary, systemd support, MIT license.

Run it in one minute

Install the binary, start the daemon, point OpenCode at it. The example config is enough to see it work.

View on GitHub Example config