krites mcp¶
What krites mcp does¶
Expose krites' commands as Model Context Protocol
tools, so an AI assistant can drive a shoot. mcp is a command group and always
takes a subcommand.
Subcommands¶
| Subcommand | What it does |
|---|---|
mcp start |
Serve over stdio — the transport most assistants expect |
mcp stream |
Serve over HTTP |
mcp tools |
Write the tool definitions to mcp-tools.json for inspection |
mcp claude |
Register krites with Claude Desktop |
mcp cursor |
Register krites with Cursor |
mcp vscode |
Register krites with VSCode |
Flags¶
| Subcommand | Flag | Default | What it does |
|---|---|---|---|
| all | --log-level |
debug, info, warn or error |
|
mcp stream |
--host |
Host to listen on | |
mcp stream |
--port |
8080 |
Port to listen on |
Plus the global flags on every command.
Unlike krites studio, mcp stream does not enforce loopback.
It has no per-launch credential and no --host validation, so binding it to a
routable address exposes the tool surface to anything that can reach the port.
Leave --host unset unless you have a specific reason not to.
Which tools an agent actually gets¶
26 tools, not one per command. Commands that produce pixels, egress an image, start a server or destroy work are excluded from the tool surface:
krites_changelog krites_ingest
krites_config_edit krites_init
krites_config_get krites_reset
krites_config_list krites_straighten
krites_config_migrate-credentials krites_update
krites_config_path krites_verdict
krites_config_set krites_version
krites_config_trust krites_xmp
krites_config_unset
krites_config_validate
krites_crop
krites_cull
krites_dedup
krites_develop
krites_docs
krites_docs_ask
krites_doctor
krites_doctor_report
Regenerate that list at any time with krites mcp tools, which writes
mcp-tools.json into the current directory.
Why export, remove, review and reset are missing¶
Deliberately. An agent must not be able to produce a deliverable, send a client's photographs to a third-party model, or delete work that cannot be rebuilt:
| Excluded | Reason |
|---|---|
export, remove, xmp write |
they produce pixels, or write into your original folder |
review |
it sends the image to a third-party model |
studio, launch |
they start a long-running server and open a browser |
reset frame, reset shoot, reset exports |
they delete work, with no confirmation prompt |
Note that krites_reset and krites_xmp do appear — as the parent command
groups. Every subcommand that would actually do something is excluded, so calling
the parent gets an agent nothing but a usage error.
What an agent can do is everything that only ever writes a reversible record: ingest a folder, cull it, list bursts, set a verdict, propose a straighten, crop or look. That is the whole point of the split — an agent can prepare work, and a human commits it.
The studio is not an MCP surface¶
krites studio and the MCP server are separate. Running mcp start does not
start the studio, and an agent connected over MCP cannot reach the studio's HTTP
API — that API has its own per-launch credential. See the
studio API reference.