hydra-mcp
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| HYDRA_DECK_URL | No | The base URL of the deck server. The MCP server connects to this address. Defaults to http://127.0.0.1:8140. | http://127.0.0.1:8140 |
| HYDRA_OUT_PORT | No | The port used by the optional video out process (out/hydra_out.py). | 8141 |
| HYDRA_DECK_HOME | No | Directory where the deck stores its configuration, live document, and state files. | ~/.hydra-deck |
| HYDRA_DECK_PORT | No | The port on which the deck server listens. Used to build the default HYDRA_DECK_URL when not explicitly set. | 8140 |
| HYDRA_DECK_TOKEN | No | If set, this token is sent as X-Hydra-Token when making requests to the deck's control API. Required only if the deck has a token configured. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
| prompts | {
"listChanged": false
} |
| resources | {
"subscribe": false,
"listChanged": false
} |
| experimental | {} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| hydra_stateA | Which Hydra tabs are connected and whether they can be trusted right now. Returns every tab with its role (exactly one Two fields are not formalities: visible false means the tab is in the background. Browsers throttle its frames, so screenshots stall and what you see may be minutes old. Ask for it to be brought to the front rather than working blind. mic 'prompt' means the browser is still waiting for someone to click Allow. Audio-reactive code (a.fft) reads zeros until they do — and nobody may be at the machine. Never assume the click happened. |
| hydra_documentA | The shared composition: ordered named blocks, each with its code and author, plus
the whole thing as one text with Read this before writing, so you extend the piece instead of talking over it — and so you can see which outputs (o0–o3) are already taken. |
| hydra_lookA | A screenshot of the primary tab's canvas, as an image, downscaled to This is the only evidence of what the visuals look like. A block that evaluated cleanly can still be black, blown out or static, so look after writing and before describing anything. It shows whatever render() currently shows — one output, or the 4-up grid. A hidden tab can't be photographed (its frames are throttled, and a minimized window's canvas has no size); the error says so. |
| hydra_referenceA | Hydra's transform functions, read live from the running engine — so it lists exactly what this tab's hydra-synth has, including anything added with setFunction.
Types, and what they chain onto: src start a chain: osc, noise, voronoi, shape, gradient, solid, src, prev coord move pixels: rotate, scale, repeat, kaleid, scroll… color change colour: color, brightness, contrast, hue, colorama, luma… combine mix another source in: add, blend, mult, diff, layer, mask combineCoord warp by another source: modulate, modulateScale, modulateRotate… Check a name here before using one you are not sure of — an invented function is refused by the dry run anyway, but this is cheaper. |
| hydra_blockA | Write one named block of Hydra code and run it, leaving every other block playing.
DRY RUN FIRST. With validate=true (the default) the code runs in a hidden second
Hydra instance before it touches the live canvas. A misspelled function comes back as
a ReferenceError; a broken shader comes back as the GPU's own compile log. Both are
refused with Human-written blocks are refused (error Things that trip people up:
Examples: osc(20, 0.1, 0.8).kaleid(4).out() noise(3).modulate(osc(10), 0.2).color(0.9, 0.3, 1).out(o1) src(o0).scale(1.01).rotate(0.01).blend(osc(8), 0.1).out(o0) shape(4, () => 0.3 + a.fft[0] * 0.4).out(o2) |
| hydra_crossfadeA | Like hydra_block, but fade from what is on screen to the new code over How: the new code is dry-run, then a copy runs into a spare output while the old chain keeps playing; a third output blends old into new and is shown; at the end the visible output switches to the new code and both borrowed outputs are blanked. Needs: one visible output (not the 4-up grid), and two outputs that no other block
draws into or reads — otherwise it refuses with Example (o0 on screen, o2 and o3 free): hydra_crossfade("base", "voronoi(8, 0.3, 0.2).color(1, 0.4, 0.8).out(o0)", seconds=6) |
| hydra_dropA | Remove one block from the composition and blank the outputs it drew into (unless another block still draws there) — Hydra keeps showing whatever an output was last given, so forgetting the code alone would leave its picture frozen on screen. Returns |
| hydra_sketchA | Replace the entire composition with a new sketch: dry-run all of it, hush, then run. Use this to start a fresh idea, not to add to one — hydra_block adds. Split the sketch into blocks with marker lines so each part can be revised on its own later: Code with no markers becomes one block called If the document holds any human-written blocks this is refused ( |
| hydra_defineA | Add a custom transform written in GLSL (Hydra's setFunction), usable in any chain afterwards like a built-in. type and what the body receives / must return:
src _st (vec2 coords) -> vec4 colour
coord _st -> vec2 coords
color _c0 (vec4 incoming colour) -> vec4
combine _c0, _c1 (two colours) -> vec4
combineCoord _st, _c0 -> vec2
It is compiled inside a throwaway chain before anything goes live, so a GLSL mistake
comes back as the compiler's own error. On success it is stored as block Example: hydra_define("stripes", "src", "return vec4(vec3(step(0.5, fract(_st.x * count))), 1.0);", inputs=[{"name": "count", "type": "float", "default": 10}]) then: stripes(20).rotate(0.3).out(o1) |
| hydra_renderA | Show one output full-screen ("o0", "o1", "o2", "o3") or all four in a grid ("all"). Every output keeps rendering either way, so a chain on o1 can feed o0 while only o0
is visible. This does not change the document — to make the choice permanent, put
|
| hydra_transportA | Set the global clock and canvas size. Pass only what you want to change. speed multiplies This changes the whole piece, human-written blocks included — say so if someone is playing along. |
| hydra_sourceA | Load an external image stream into s0–s3, then read it in any chain with src(s0). kind: "image" target = an http(s) URL. It must allow cross-origin use (CORS) or WebGL cannot read it — the error says so. "video" target = an http(s) URL, same CORS rule; loops muted. "cam" target = camera index ("0"). The browser asks for permission: expect state "pending_permission" until someone at the machine clicks Allow, and check hydra_state rather than assuming. "screen" screen capture. Browsers only allow it from a click inside the page, so a remote call usually cannot start it; the error explains. "clear" release the slot. Returns Then: src(s0).kaleid(4).out(o1) |
| hydra_audioA | Check and tune audio reactivity (Hydra's With no arguments: the current a.fft band values, volume and microphone permission — a quick way to see whether sound is getting in at all. enable=true Hydra only listens if it was started with the microphone, so this reloads the deck page with ?audio=1. The piece is replayed afterwards; the browser asks for the mic, and a.fft reads zeros until someone clicks Allow. Video out has to be started again after the reload. bins number of frequency bands in a.fft (1–64; default 4). smooth 0–1: how slowly band values fall (higher = smoother). cutoff noise floor subtracted before scaling. scale how much raw loudness it takes to reach 1. show true draws the band meters on the canvas while you tune; false hides them. Use it in code inside a function, or it is read once and frozen: shape(4, () => 0.2 + a.fft[0] * 0.5).out(o2) |
| hydra_video_outA | Send the primary tab's canvas out of the browser as a native video source: a Syphon server named "Hydra" (macOS). TouchDesigner receives it with a Syphon Spout In TOP. TouchDesigner names Syphon senders App:Server, and the publisher runs as Python, so the sender to pick is "Python:Hydra" — plain "Hydra" matches nothing and leaves the TOP on its checkerboard. With hydra_out.py started with --ndi, the same
frames also go out as an NDI source, which receivers list as " (Hydra)". NDI
reaches other machines on the network; action: "start", "stop" or "status". fps: 1–60, the capture rate to request. Needs out/hydra_out.py running — a separate process with its own Python 3.12, because
the Syphon library only supports up to 3.12 (see the README). Start waits two
seconds, then returns both ends:
page_after_2s sent_fps, dropped — what the tab actually sent
out_after_2s fps, syphon_clients — what was published, and whether anything is
receiving
Report those numbers, not the fps you asked for. Frames are dropped rather than queued
when the receiver lags, so a high Chromium tabs only (MediaStreamTrackProcessor). The primary must stay visible: a hidden tab stops producing frames. |
| hydra_hushA | The panic button: blank every output, clear update(), reset speed to 1, and empty the document — every block, human-written ones included, in every connected tab. The document is archived first (the path comes back as |
| hydra_saveA | Save the current composition — every block, with its author — to the sketch library (hydra-mcp/sketches/.json), plus a 480-px thumbnail of what is on screen. name: letters, digits, _ and -, up to 60. tags: a few words to find it by later ("feedback", "audio", "slow"). An existing name is refused unless overwrite=true. |
| hydra_loadA | Replace the composition with a saved sketch — the same dry run, hush and run as hydra_sketch, keeping each block's original author. Refused if the current document holds human-written blocks, unless take=true. Save the current piece first if it is worth keeping. An unknown name returns the list of saved sketches. |
| hydra_sketchesA | List saved sketches: name, tags, when saved, block names, and the thumbnail path.
|
| hydra_share_urlA | A hydra.ojack.xyz link that opens the current piece in the official Hydra editor. Built here from the document text, in the editor's own ?code= format (base64 of the URI-encoded code; verified by loading such a link and reading the editor back). When the primary is an existing Hydra tab, its own editor text is used instead. Nothing is uploaded or posted to Hydra's gallery — the code travels only inside the link, so very long pieces make very long links. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 19 tools
Most tools map cleanly to a distinct action in the Hydra workflow, and the descriptions are unusually precise. A few close pairs exist — hydra_sketch vs hydra_load both replace the whole composition, and hydra_block vs hydra_crossfade are write operations differing only in transition — but their purposes are clearly separated in the docs.
All tools share the hydra_ prefix and lowercase snake_case, which makes the set feel cohesive. However, the names mix nouns (hydra_state, hydra_document, hydra_reference) with verbs (hydra_drop, hydra_render, hydra_save), so the pattern is predictable but not a strict verb_noun convention.
At 19 tools, the server is on the heavy side of the ideal range. The count is defensible given the broad domain — composition editing, state inspection, rendering, media sources, audio, persistence, and video output — but a few tools could plausibly be consolidated without losing clarity.
The core composition lifecycle is well covered: read, create, replace, fade, delete, panic, save, load, list, and share. Minor gaps remain, such as no direct way to delete a saved sketch, reorder blocks, or undo changes, but an agent can accomplish the main Hydra workflows without hitting dead ends.