gemini-mcp

MCP server for Google Gemini media generation. Exposes thirteen tools to Claude over stdio: list available models, generate/edit/compose images, generate a consistent set of images from a master prompt, multi-turn image refinement (Interactions API), video generation (omni), music generation (Lyria), an async result poll for long generations, and Files API upload/list/delete for reusable image references. Output is written to disk by default (path returned) or returned inline as base64. Built on the Gemini v1beta API (generativelanguage.googleapis.com) using the Nano Banana / Nano Banana Pro (images), omni (video), and Lyria (music) model families.

Developed and maintained by AI (Claude Code).

Environment Variables

Confirmations

Sending a local file to Google (the images / master_images / video_path inputs, or gemini_upload_file's path) and every delete ask for confirmation first. A client that can show a confirmation prompt (Claude Code) gets the real prompt, unless MCP_CONFIRM_ELICITATION=off. Elsewhere the first call does nothing and returns a preview (the resolved local paths with their MIME types and sizes, or the method/path being deleted) plus a confirmToken; only a repeat call with the same arguments and that token proceeds. The token is single-use, expires, and is bound to the exact arguments — a changed prompt or a swapped file between the two calls is refused. Pure text prompts, URLs, files/ references and base64 inputs are not gated.

Long generations and client timeouts

4K / Pro-model generations can outrun an MCP host's own tools/call timeout (error -32001). The server sends notifications/progress heartbeats so hosts that reset their timeout on progress wait it out. If the host still gives up, the server-side generation usually completes anyway: the image is written to the output dir, gemini_interact also writes an <image>.json sidecar recording the interaction_id, and continue_last: true resumes the interaction the lost response belonged to.

When a chained call 404s

A 404 on a request carrying previous_interaction_id is not proof the chain expired. The only 404 body observed live is generic — "Requested entity was not found." — and never names which entity. An unknown or renamed model id, and an expired Files API `files/…` uri (~48h TTL), return exactly the same thing. So the server no longer asserts a cause it can't establish: the upstream text is surfaced verbatim, and gemini_interact runs an experiment to find out which it was.

Most often the id isn't missing at all — it just isn't visible yet. The interactions store is eventually consistent, and a freshly created id can 404 while the same id resolves fine minutes later; heavy turns (4K, Pro, thinking_level: high) are the likeliest to hit it, which is exactly the turn you most want to chain from. So a chained 404 is retried with exponential backoff for up to 120s (GEMINI_CHAIN_RETRY_MS) before anything is declared broken. The 404 generates nothing and isn't billed, so the wait costs only time.

After that budget is spent, the tool looks up that id's sidecar, re-attaches the image it produced, and re-issues the request without the chain:

• The re-issue succeeds → the chain really was the problem, and you get your image anyway,

reported as chain_recovered: { expired_interaction_id, reanchored_on }. The 404'd attempt generates nothing, so this costs the one generation you'd have paid for re-anchoring manually.

• The re-issue 404s too → the interaction id was never the cause. You get told exactly that,

with the upstream text, and pointed at the model id and any files/… uri instead of being sent to chase an interaction that was fine all along.

• No sidecar matches the dead id → the original error, rather than a guess. Re-anchoring on

the wrong picture would silently corrupt the edit.

Separately, continue_last no longer dies with the server process: with no in-memory id it resumes from the newest <image>.json sidecar in the output dir and reports continued_from_sidecar: true. That case was never an expired chain at all — the interaction was alive upstream the whole time; only our memory of its id was gone.

For hosts whose timeout can't be tamed (e.g. Claude Desktop, a fixed ~30s cap that ignores progress), two guards make re-issuing safe and unnecessary:

• `async: true` returns a job_id immediately instead of the image, so the call can't

time out at all; poll gemini_get_result with the job_id until it's done.

• `idempotency_key` makes a repeat call idempotent — a retry with the same key returns the

recorded result (reused: true) instead of billing a second generation. (Even without a key, two identical in-flight calls are deduplicated automatically.)

Tools

Generation tools also share three throughput/latency controls: async: true (return a job_id immediately), max_wait_ms (wait up to a budget, then hand back the job_id — fast results stay in-band, slow batches never trip the host timeout), and idempotency_key (a retry returns the recorded result instead of re-billing). On a hosted deployment, a gemini_image_set result with more than one image also carries a `bundle_url` — one signed URL for a zip of every image in the set. Like every hosted link it lives at most 24h (the host's blob-store ceiling); gemini_sign_media re-mints an expired one. (A set too large to zip safely in memory skips the bundle and says so via bundle_skipped; the per-image links are unaffected.)

Seeing your images (hosted)

On a hosted deployment there is no filesystem, so a generated image has to come back as something you can open. It does: every result includes a URL, with no configuration.

{
  "images": ["https://mcp.nullnet.app/b/<account>/gemini/gen/2026-07-29/ab12cd34-a-cat.png?exp=…&sig=…"],
  "media":  [{ "url": "https://…", "r2_key": "gen/2026-07-29/ab12cd34-a-cat.png",
               "expires_at": "2026-07-31T12:00:00.000Z",
               "curl_hint": "curl -sS -o a-cat.png \"https://…\"" }]
}

Those links need no auth header — the signature is in the URL — so they work in a browser, in curl, and in a chat message. They expire after at most 24h — the host's blob store refuses a longer signature — and the objects behind them are swept on the host's retention schedule. The r2_key is the durable handle for that window:

• `gemini_sign_media` (hosted deployments only) mints a fresh signed URL from an r2_key,

so an expired link never forces you to re-generate — and re-pay for — the image.

• The interaction id is stored beside each image, so a hosted deployment recovers a lost

turn the way the local one does: continue_last: true survives a restart, a chained 404 is re-anchored on the image that interaction produced, and gemini_list_recent_media shows the prompt, model and interaction id for every stored object rather than a wall of keys.

• `gemini_view_media` (hosted deployments only) returns the image itself, inline, from an

r2_key. A link is not something a model can look at, so without this a generate → refine loop runs blind. It is a separate call rather than bytes on every result: you pay image tokens only on the turns you actually look.

• `gemini_upload_file` with `r2_key` turns media this server generated into a Files API

reference (the server reads it back with its own key — no signature for you to mint), so a generated image can become the reference image for the next generation in one cheap call.

• Idempotent replays (idempotency_key) re-mint the URLs inside the recorded result before

returning it, so a reused result never carries a dead link.

Why this matters: MCP's inline image content blocks (inline: true) are visible to the assistant but many chat clients never render them to the user, and the assistant cannot extract bytes back out of its own context to save them elsewhere. A generation could bill successfully and be invisible. A URL is the portable answer; inline remains available, but it is no longer the only way to receive media.

How the bytes are served

The host stores each generated object and serves it back at a signed, expiring URL — https://<host>/b/<account>/gemini/<key>?exp=&sig=. No setup, and no auth header: the signature in the link is the authorization, so curl and a browser both work.

Auth is a signed, expiring URL rather than an unlisted key. Random keys would be simpler, but they never expire and never revoke: anything that ever logged or forwarded the link keeps working forever. A signature scopes access to one object with a deadline, and rotating the host's signing key (MCP_BLOB_SIGNING_KEY, which mcp-host injects together with MCP_BLOB_BASE_URL; this server never chooses either) invalidates every outstanding link at once. The tradeoff is that links are long and cannot be shortened by hand.

For assistants relaying a result

Show the user the URL. If your sandbox has network egress to the host, fetching it and attaching the bytes as a file gives the nicest result; otherwise present the link itself. Whether a given client renders ![](url) markdown inline varies by client — a bare URL is the safe form, and a markdown link is a reasonable enhancement where you know it renders.

Retention

Retention is the host's, not a setting of this server. On mcp-host, the blob store prunes generated media (gen/) and signed uploads (up/) on its own schedule (see mcp-host docs/BLOB-STORE.md; default 30 days), while the character/style library (lib/) is written as retention-exempt, so saved entries never expire. Every signed link is clamped to the store's 24h ceiling, so it never outlives what the store will honour; gemini_sign_media re-mints one for as long as the object exists. The variables this server reads for it are injected by the host: MCP_BLOB_BASE_URL and MCP_BLOB_SIGNING_KEY.

Prompts are stored too. Each generated object on a hosted deployment gets a small JSON record beside it holding the prompt (first 500 characters), model and interaction id — that is what gemini_list_recent_media shows and what lets a lost turn be recovered. The record lives and is swept with its object, so it is kept for the same retention window, not longer. Locally, gemini_interact writes an <image>.json sidecar (model, interaction id, options and any text the model replied with — not the prompt itself) next to each image in the output dir; those files are yours to keep or delete and never expire on their own. If your prompts describe real people, bear in mind they sit in your bucket for the retention window.

Sending reference images without burning context

Every tool that takes a reference image — gemini_image_generate, gemini_image_edit, gemini_image_set, gemini_interact, plus gemini_video_generate (reference stills) and gemini_music_generate — accepts them four ways. Only one of them costs model context:

images_base64 is the fallback of last resort. It costs roughly 14k tokens per modest photo, and it is silently corrupted whenever the file read that produced it was truncated — the payload still looks like base64, so the failure surfaces as a bad generation rather than an error. Prefer any of the other three.

images_url — the server fetches it

{ "prompt": "make it look like winter", "images_url": ["https://example.com/photo.jpg"] }

Fetches are restricted to public https:// URLs — private, loopback and link-local hosts are refused (IPv6 literals are parsed, so [::ffff:7f00:1] is caught as loopback), every redirect hop is revalidated, and each hop is bounded by a timeout. The response must be Content-Type: image/* and is capped at 15MB, enforced while streaming rather than trusted from Content-Length. A failure names the offending URL. Anything over 6MB is uploaded to the Files API and referenced by uri instead of inlined, since generateContent caps a whole request near 20MB.

images_file_uris — upload once, reference many times

// 1. upload
{ "tool": "gemini_upload_file", "url": "https://example.com/photo.jpg" }
// → { "file_uri": "files/abc123", "mime_type": "image/jpeg", "expires": "..." }

// 2. reference it, as many times as you like
{ "prompt": "make it winter",  "images_file_uris": ["files/abc123"] }
{ "prompt": "make it sunrise", "images_file_uris": ["files/abc123"] }

Uploads are retained ~48h; after that the reference stops resolving (as a generic 404 — see the chained-404 section above). gemini_image_set fetches or resolves such a reference once and passes it to the master and every scene call.

On stdio builds, a local images path that gets referenced more than once in a session is uploaded to the Files API automatically (keyed on path + mtime + size), so repeated edits of the same photo stop re-sending the bytes. Editing the file invalidates the cached upload.

Signed upload URLs — no token at all (hosted)

This is the intended path for an agent with a shell: disk file → curl → r2_key → tool call, with the image never entering the conversation. gemini_get_upload_url mints a short-lived (~10 min) signed PUT URL, and the shell uploads with zero auth headers — the signature in the URL is the authorization, mirroring how the download links work.

# 1. tool call: gemini_get_upload_url { filename: "photo.jpg", content_type: "image/jpeg" }
#    → { upload_url, r2_key, expires_at, curl_hint }

# 2. shell:
curl -sS -X PUT -H "Content-Type: image/jpeg" --data-binary @photo.jpg "$UPLOAD_URL"
# → { "r2_key": "up/<tenant>/2026-07-31/ab12cd34-photo.jpg", "size_bytes": 812345, ... }

The signature covers one tenant-scoped object key, the declared content type and the expiry; uploads are capped at 15MB, enforced while reading the stream. Only raster image types are accepted (jpeg/png/webp/gif/avif/heic/heif/bmp/tiff) — SVG is deliberately refused, because an SVG is a scriptable document and /media serves from the server's own origin. The returned r2_key is then usable three ways: directly as images_r2_keys on any generation tool (the server reads its own bucket — no bytes in the conversation), permanently via gemini_save_character, or as a ~48h Files API reference via gemini_upload_file({ r2_key }). Uploads themselves follow the host's blob-store retention schedule (up/ prefix).

Character & style library (hosted)

Recurring subjects and styles can be saved once, per account, with no expiry (the library's lib/ objects are written retention-exempt, so the host's prune skips them):

// once:
{ "tool": "gemini_save_character", "name": "finn",
  "description": "6-year-old boy, curly red hair", "image_r2_key": "up/…/photo.jpg" }
{ "tool": "gemini_save_style", "name": "bold-cartoon-sports",
  "prompt_fragment": "bold cartoon style, thick outlines, saturated colors" }

// afterwards, on any generation:
{ "tool": "gemini_image_set",
  "master_prompt": "Finn on a soccer field",
  "scenes": ["kicking the ball", "celebrating a goal"],
  "characters": ["finn"], "style": "bold-cartoon-sports" }

Naming a character attaches its saved reference image and weaves its description into the prompt; naming a style appends its fragment (and attaches its reference image, if it has one). gemini_image_set passes character references to the master and every scene call, which is what keeps the subject consistent across the set.

Quick Start

{
  "mcpServers": {
    "gemini": {
      "command": "npx",
      "args": ["-y", "@chrischall/gemini-mcp"],
      "env": {
        "GEMINI_API_KEY": "your-api-key-here"
      }
    }
  }
}

See SKILL.md for full usage documentation.