AI Media Router

Add assets to a collection, creating it when missing.

Append assets to a reference, optionally with a role (e.g. "side").

Cancel a waiting or running job: it becomes failed with `cancelled`, and jobs waiting on it fail too. The provider is not asked to stop, so a running generation may still bill there.

Check a spec without saving or calling the provider: validates it (pointers, template variables), replays recorded fixtures, and with model/input renders the exact HTTP requests it would send.

Join two or more video assets in order into one MP4 (first clip's size, 30 fps); returns the new asset.

Delete an asset's file and library entry; it leaves every reference and collection. Jobs keep their record of it.

Delete a budget; its jobs stay in the history.

Delete a recipe; jobs made from it stay in the history.

Delete a reference; its assets stay in the library.

Delete a snippet; inputs that still use it fail with the list of known snippets.

Check the installation: data directory, database, ffmpeg/ffprobe, which providers have keys, spec override errors, recorded fixtures of configured providers, jobs no worker is polling, and the last day's failure kinds. online=true also pings every configured provider. Run it first when something does not work.

Estimate a generation's USD cost: an offer's price (offer_id from search_prices) times the duration or image count read from input, plus the median of past jobs on this provider and model. Says which formula and input fields it used. With recipe (and vars) instead: every step of that recipe and the sum.

Export a collection as a ZIP in <home>/exports/: the files, their previews, and collection.json with each asset's metadata (review status, rating, notes), the job that made it (provider, model, input, cost) and its references. only_approved=true ships only assets approved with review_asset.

Save one frame of a video as a PNG asset: at "first", "last" or seconds. Use the last frame as the start image of the next clip.

Start a generation. Returns the job immediately (status running); poll it with get_job. Results are downloaded into the library. Library files go into input as "{{asset:<id>}}" / "{{ref:<name>}}"; save_to files the outputs into a collection and/or reference. `recipe` names a saved single-generation preset (list_recipes shows the user's presets), so provider and model can be omitted.

Start up to 100 jobs at once: items (e.g. one prompt on several provider/model pairs) x matrix (top-level input fields to vary). Item input merges over defaults.input. With defaults.budget the sum of estimates is admitted first. Returns the batch; follow it with get_batch.

Metadata and local path of one asset.

A batch's jobs per status, each job's outputs, and every output file. wait_secs blocks until all finish (max 50).

Job status, provider progress and stored result files. wait_secs blocks until done (max 50); debug=true adds the raw provider exchanges.

Generation graph of an asset: the assets and references it was made from (ancestors) and everything made from it (descendants), with the job of each step.

How to call one provider: model id format, an example generate call, docs link, upload support, known quirks.

One recipe by name.

One reference by name or id, with its assets in order and their local paths.

Read a provider's active spec (TOML), whether it is built-in or a local override, and its fixture names. First step of fixing a provider.

Per provider and model: jobs, done/failed share, failure kinds, average rating, approved share, average USD cost and seconds to finish. Filter by provider, model or collection. Use it to pick a model that worked before.

Copy a local file into the asset library, e.g. a reference image. Files made by ai-media-router carry their provider, model, input, job and parent assets: import restores prompt, tags and lineage, and returns that provenance.

Budgets with their limit, spent and remaining USD.

List collections (projects) with their asset counts. search_assets {collection} lists one's assets.

Generation history, newest first, with provider, model, input, cost, input assets and outputs. Filter by reference, collection, status or error_kind.

Every provider key any loaded spec uses: whether it is set, where it comes from, and the command that sets it. Never returns key values (only the last 4 characters), so it is safe to show the user; keys are added by the user in a terminal.

List providers, whether their API keys are set, and whether each runs the built-in spec or a local override.

List recipes with their provider, model, input or steps, and variables.

List references, most recently changed first; filter by kind or a substring of name, description or tags.

List snippets with their text.

Put an audio track (audio asset or another video's sound) under a video, replacing its sound; returns the new asset.

Draw an image (packshot, logo) over an image or video at x, y pixels, optionally scaled to width; returns the new asset.

Dry run of generate with the same arguments: checks the provider and keys, snippets and placeholders (assets, references, parent jobs), upload support, save_to, cost and budget, and renders the exact HTTP requests. Sends nothing, spends nothing; `ready` says whether generate would start. Use it before an expensive or unfamiliar generation. With recipe and vars instead: every step of run_recipe (checks named <step>/<check>), the summed estimate and the budget.

Change an image's or video's aspect ratio (e.g. 16:9 to 9:16) by centered crop or black padding; returns the new asset.

Remove assets from a collection (not from the library); an emptied collection disappears.

Remove assets from a reference; they stay in the library.

Delete a provider's local override so the built-in spec is used again.

Review assets: status approved/rejected/draft, rating 1-5, notes. Rejected assets are skipped when a {{ref:…}} expands; export_collection {only_approved} ships only approved ones; get_stats averages ratings per model.

Start a generation from a recipe with values for its variables (defaults fill the rest). Returns the job like generate. A recipe with steps starts every step at once in one batch, admitting the sum of their estimates against `budget`; it returns the last step's job, and get_batch {batch_id} follows the whole run.

Save a generation template: from_job copies a working job, or give provider, model and input. "{{var:<name>}}" in input marks what changes per run; every recipe also appears as the MCP prompt recipe:<name>. A chain: `steps` (generations and ffmpeg edits linked by "{{step:<name>}}"), or from_job / from_asset with chain: true to take the final job's whole lineage.