MCP server de camera
Commit 14 nov 2026Un MCP server y monitor de vida silvestre que utiliza detección de movimiento y modelos de IA para identificar animales, personas y vehículos desde transmisiones de cámaras IP domésticas.
También conocido como ecoop/camera-mcp
- Stars
- 0
- Herramientas
- 1
- Descargas, 30 días
- —
- Última versión
- —
- Último commit
- 14 nov 2026
- Licencia
- MIT
¿Qué es camera MCP?
Un MCP server y monitor de vida silvestre que utiliza detección de movimiento y modelos de IA para identificar animales, personas y vehículos desde transmisiones de cámaras IP domésticas. Expone 1 herramienta, como list_cameras.
Resumen generado a partir del README del proyecto
Acerca del MCP server de camera
El monitoreo de la vida silvestre y la seguridad del hogar a través de transmisiones de cámaras IP es lo que camera-mcp permite a propietarios y entusiastas de la naturaleza. Este sistema utiliza un enfoque basado en el movimiento para vigilar las transmisiones de las cámaras, activando modelos de IA como MegaDetector y SpeciesNet solo cuando se detecta movimiento para ahorrar recursos computacionales. Identifica animales, personas y vehículos, registrando cada llegada y salida como una visita en una base de datos SQLite local con sus respectivas miniaturas y clips de video.
Los usuarios interactúan con sus cámaras a través de un cliente MCP como Claude, utilizando herramientas para listar cámaras, tomar capturas o consultar el historial de actividad. El servidor puede proporcionar resúmenes del conteo de especies o detalles de visitas específicas basadas en rangos de tiempo o grupos de animales. La configuración requiere Python 3.12 o 3.13 y ffmpeg, con las credenciales de la cámara gestionadas de forma segura mediante el Keychain de macOS o archivos de credenciales locales. Está diseñado para ejecutarse en una máquina local, como un Mac Studio, para procesar transmisiones de video sin saturar la red o el hardware.
Qué puedes hacer con él
- 1Consultar la actividad de vida silvestre durante la nochePregúntale a la IA qué animales visitaron el jardín durante la noche.
- 2Verificar el conteo de especies específicasSolicita un resumen de cuántos pájaros o mapaches se han visto recientemente.
- 3Verificar la cobertura y el estado de la cámaraDetermina si una cámara específica está activa actualmente y vigilando si hay movimiento.
- 4Recuperar evidencia visual de una visitaObtén la mejor imagen fija de un avistamiento de animal registrado.
Escrito a partir del README del proyecto. Cada caso de uso se verifica con una cita del mismo.
Herramientas de camera MCP 1
Detectado desde el READMElist_cameras
| list_cameras | — |
Configuración
ANTHROPIC_API_KEYASK_DAILY_USD
| Nombre | Descripción | Tipo |
|---|---|---|
| ANTHROPIC_API_KEY | — | |
| ASK_DAILY_USD | — |
Verificaciones automáticas
Ejecutado 14 nov 2026Informe completo →Aún no verificado. Un servidor necesita una puntuación de 80, ningún hallazgo grave y suficientes datos para que podamos revisarlo. Cómo funcionan las verificaciones · Qué corregir para obtener más puntos
- Repositorio de origen vinculado y accesibleecoop/camera-mcp is public and reachablegithub.com/ecoop/camera-mcp ↗Aprobado
- El paquete indica este repositorio como su origenNo verificado
- Procedencia de la compilación (atestación de Sigstore)No verificado
- Licencia de código abierto (OSI)MITspdx.org/licenses/MIT.html ↗Aprobado
- Commit o lanzamiento en el último añoLast commit or release 2026-11-12github.com/ecoop/camera-mcp ↗Aprobado
- OpenSSF ScorecardNo verificado
- El nombre del paquete no es similar a un paquete popularNo verificado
- No hay avisos críticos en el paquete ni en sus dependenciasNo verificado
- Sin instrucciones ocultas para el modelo1 tool description and the README read; no override phrases, credential paths or hidden charactersAprobado
- Las herramientas destructivas están etiquetadasNo tool named like delete, drop or execAprobado
- El endpoint remoto responde al handshake de MCPNo verificado
README
Ver en GitHubcamera-mcp
Last updated: 2026-11-12
An MCP server and an always-on monitor for home IP cameras, built to watch for wildlife on one machine without saturating it.
- Motion first, models second. A lightweight motion frontend watches every camera on small grayscale frames and opens a window only when something moves; only those windows reach the detector. Analysis regions, motion masks and exclusion zones cut the work further, so six cameras fit on one Apple-Silicon GPU.
- Detect, then identify. MegaDetector finds animals, people and vehicles; SpeciesNet names the animal, limited to species found in your area. Each arrival-to-departure is logged as one visit in SQLite, with a thumbnail, a full-resolution still and a clip.
- Junk control. A memory of fixed scenery, "ignore" zones that drop boxes before they are classified, and a reversible gate for unidentified visits keep leaves, fruit and feeder roofs out of the log.
- Ask Claude. The MCP tools below answer live and history questions (what came by overnight, is the yard clear) and say plainly when a camera wasn't being watched.
- A web dashboard for browsing visits and analytics, tagging detections down to single boxes, favourites, a labels map, drawing zones, and a plain-English ask box.
- A morning brief by email or iMessage.
The design is in docs/phase2-design.md, backed by the SpeciesNet eval. SpeciesNet's full label list is in docs/speciesnet-labels-v4.0.3a.txt.
Tools
| Tool | What it returns |
|---|---|
list_cameras |
name, label, tags, aliases, and whether each camera answers right now (a TCP check, no decode) |
snapshot(camera, max_width=1024) |
one current frame as a JPEG image, downscaled to max_width; from the low-res stream if the camera's main stream is busy (Tapo cameras allow only a couple of main-stream clients) |
find_visits(since, until, cameras, tag, taxon, overlapping, min_score, limit) |
animal visits in a stretch of time, newest first; still_there marks one happening now |
activity_summary(since, until, group_by, cameras, tag, taxon) |
visit counts by species, group, camera, local hour or day, following your tags |
visit_media(visit_id) |
the thumbnail of a visit's best sighting |
monitor_status() |
per camera: watched now?, last frame and motion, looks moved?, visits in progress, errors |
Resources: cameras://list (JSON) and cameras://{name}/latest.jpg.
All tools are read-only. Cameras are addressed by name (loosely: "west alley" finds west_alley);
a source URL or credential is never returned. The history tools read the monitor's
data/monitor.db read-only.
since and until take ISO times or durations ago (15m, 4h, 3d, 2w), today,
yesterday. taxon takes a group (songbirds, jays, crows, doves/pigeons, hawks, owls,
cat, dog, rat, squirrel, skunk, raccoon, opossum, fox, coyote, person), the union
night-critters, a class (bird, mammal) or a common name. overlapping finds visits that
overlapped another kind on the same camera: taxon=cat, overlapping=bird is cats among the birds.
Every history answer carries coverage: the share of the time each camera was actually being
watched (null if the log predates that record). "Nothing seen" only means nothing was there when
coverage is near 1. "Is the dog yard clear?" is find_visits(since="15m", tag="dog_yard", taxon="night-critters").
Setup
Needs Python 3.12+ and the ffmpeg binary (Homebrew: brew install ffmpeg). Species identification
(SpeciesNet, the detect extra) needs Python below 3.14 until SpeciesNet supports it; everything
else also runs on 3.14.
uv venv .venv && uv pip install --python .venv/bin/python -e '.[dev]'
cp cameras.example.yaml cameras.yaml
cameras.yaml is git-ignored. A Tapo camera needs only a name and its IP:
cameras:
- name: east_fence
label: Back fence, east
host: 192.0.2.10
The RTSP URL is built from the host (rtsp://<user>:<pass>@<host>/stream1). The login is the
Camera Account set in the Tapo app, and it never goes in a file. On macOS, store it once in the
Keychain; security asks for the password, so it isn't typed on the command line:
security add-generic-password -U -s camera-mcp -a <camera user> -w
$CAMERA_USER and $CAMERA_PASS, if set, take precedence (for other systems, and the tests). On a
headless or remote host the login Keychain is locked in an SSH session, so put
CAMERA_USER= and CAMERA_PASS= lines in ~/.config/camera-mcp/credentials (chmod 600,
or $CAMERA_MCP_CREDENTIALS to point elsewhere) and that is used instead. The same file also
holds the optional email and alert settings (SMTP_*, SMS_ALERT_TO, WEB_URL); see
credentials.example for every key the code reads, with placeholder
values and the exact KEY=value format it expects.
Anything else goes in source instead of host: a video file (which is how the tests run, and
how to try the server away from home) or a full URL for a non-Tapo camera, where ${VAR} is
expanded from the environment.
Use from Claude Code
As a plugin (the easy way): this repository is also a Claude Code plugin marketplace. The plugin
brings the MCP server and a Skill, camera-history, that teaches Claude how to answer from it:
common questions with the exact tool calls, and how to read coverage, paused cameras, hidden
visits and your tags honestly. It also adds two shortcuts, /camera-mcp:overnight and
/camera-mcp:yard-check [camera tag].
/plugin marketplace add ecoop/camera-mcp
/plugin install camera-mcp@camera-mcp
On install it asks for two settings: the monitor host (user@host, reached over key-based SSH;
leave it empty if Claude Code runs on the monitor's machine) and the camera-mcp folder there
(default ~/projects/camera-mcp). CAMERA_MCP_HOST / CAMERA_MCP_DIR in the environment override
them. The plugin's tools are named mcp__plugin_camera-mcp_cameras__<tool>; if you added the server
by hand before (below), remove that entry (claude mcp remove cameras) so Claude doesn't see every
tool twice, and update any permission rules that named the old tools. To try it without installing:
claude --plugin-dir plugin. The plugin lives in plugin/; tests/test_plugin.py keeps its Skill
true to the server's tools and the taxon list (plugin/skills/camera-history/reference.md is
generated from taxa.py).
By hand:
claude mcp add cameras -- ~/projects/camera-mcp/.venv/bin/camera-mcp --config ~/projects/camera-mcp/cameras.yaml --data-dir ~/projects/camera-mcp/data
No password goes into Claude's settings: the server reads the login from the Keychain, and only
snapshot needs it. Without a login the server still starts, the history tools work, and
snapshot says how to add one.
Probe the cameras
.venv/bin/python -m camera_mcp.probe
For every camera in cameras.yaml, this checks that it answers and accepts the login, and reports
what its streams deliver (codec, resolution, frame rate). Tapo cameras are checked on both
/stream1 and /stream2. It's read-only: each stream is opened, its header read, and closed.
It exits non-zero if any stream fails.
The monitor
To run the monitor as an always-on service (its real home is the Mac Studio, where the GPU keeps
SpeciesNet fast enough), see deploy/README.md. Health checks run beside it
(camera-watchdog): a text when a camera's video stops inside its watch hours, and a midday report
on whether the detector kept up.
uv pip install --python .venv/bin/python -e '.[detect]' # SpeciesNet, for animal identification
.venv/bin/camera-monitor # all enabled cameras, until Ctrl-C
.venv/bin/camera-monitor --camera "west alley" --duration 300
For each camera, the monitor:
- records the main stream into a ring buffer of 2-second files, copied rather than re-encoded;
- watches the low-resolution stream for motion, at 5 frames per second;
- identifies what's there: while there's motion, or an animal is still in view, it runs
SpeciesNet on one frame per second from the buffer.
Every box is classified, and on
wide_viewcameras the detector also looks at tiles of the frame, so small animals far away aren't lost; - logs visits: one per kind of animal per camera, from arrival until it's gone 30 seconds
(a flock is one visit with a head count; a cat among the birds is a second visit). Each visit
gets a label by vote across its frames, a group (
songbirds,jays,cat,raccoon...), a thumbnail of the best sighting and a clip with the seconds before and after.
Everything goes to data/ (git-ignored; --data-dir or $CAMERA_MCP_DATA to move it):
monitor.db (visits, their frames, motion windows, camera status), media/ (thumbnails) and
clips/. Dropped streams are reopened with backoff. The first detection loads the model, which
downloads its weights (about 500 MB) on first use.
Without SpeciesNet installed, or with --detector none, the monitor watches for motion only and
keeps a clip of every motion window. With it, clips are kept for visits, so leaves, shadows and
insects in the infrared light don't fill the disk; set training_capture: true on a camera to
keep motion clips too. On a slow machine analysis thins to one frame per segment when it falls
behind (counted as thinned in analysis_health, on the dashboard's health line and as a faint band
on the Analytics lag charts); SpeciesNet takes about 4 s per frame on an Intel MacBook Pro's CPU.
Where the cameras are goes under site: in cameras.yaml. SpeciesNet drops species not found
there, and an allowlist turns any other species into its class ("mammal", "bird"). timezone
sets the hours of the day that activity_summary reports (default: this machine's):
site:
country: USA
admin1_region: CA
timezone: America/Los_Angeles
allowlist: [domestic cat, domestic dog, northern raccoon, striped skunk, virginia opossum, aves]
The web dashboard's species filter is built from what's actually been seen (most frequent first,
with night-critters once one has shown up). To pin it to a fixed menu instead, set
menu_species: [night-critters, cat, raccoon, …] under site: — then only those show, in that
order.
The time filter takes "last 2h" … "30 days", or from–to… for a range between two local times (an empty "to" means now); every view uses it. The feeds grid lists the window's newest 60 visits, with "N of M visits · load more" underneath for the next 60.
Settings, site-wide under monitor: or per camera: pre_seconds, post_seconds,
quiet_seconds, motion_fps, pixel_threshold, min_changed, detect_every_seconds,
visit_close_seconds, min_box_conf, max_boxes, save_people_images (default true),
detect_vehicles (default false; for a driveway camera) and training_capture.
Learned scenery. Fixed objects that the detector keeps mistaking for an animal or a person
(an empty feeder, a cloth on a rail, clothespins on a line) are learned after 5 sightings over 10
minutes at the same place and size, and ignored from then on (data/scenery/). Named animals are
never ignored, however still they sit.
Unnamed-visit gate. Swaying leaves and fruit don't sit still enough to be learned as scenery,
so they become visits of a bare "animal". Per camera, unnamed_min_conf hides such a visit when
its best detector box is below that confidence, and unnamed_min_travel hides it when its boxes
moved less than that many box sizes (a leaf sways in place; an animal crosses the frame) unless
its best box reached unnamed_travel_conf (default 0.5). Both are off (0) by default. Hidden visits
are kept, with the reason, and counted as hidden in history answers;
find_visits(include_suppressed=true) lists them. Visits with a name are never hidden. Try values
against the camera's own history first (read-only):
.venv/bin/camera-backtest --since 7d --camera east_fence --min-conf 0.3 --min-travel 0.5
To check the gate by eye, pick species "animal" and tick + hidden on the dashboard: hidden
visits show a badge with the reason. Open one and mark it real or false; the viewer moves on to
the next visit in the queue picked beside next › (unlabelled in the grid, most useful, or boxes to
tag; the choice is remembered), or press next ›. Wrong sp. moves on once you've said
what it was. The backtest's lbl_real column then
counts real animals the rules would hide. A label is per visit, but a visit you mark real
can still have a junk box or two (a bird on the feeder plus a box on the clothesline), so single
boxes can be tagged too, and any tag can say what the animal actually was (a species: the right
name for a wrong-species visit, or a name for an unnamed one; never on a false one). A box with no
tag of its own takes its visit's where that must hold: every box of a false or wrong-species visit,
and of a real visit whose boxes all stayed in one place. A real visit whose boxes moved around
leaves them untagged.
In the viewer, one set of tag buttons tags the target picked in the tag row: the whole visit, or one box (pick its chip, or click it on the still: a click, not a drag). An untagged visit with several boxes starts on its most confident box, and each tag moves to the next box. Once every box on the still has its own tag, they tag the visit (noted "from boxes"): false if all are false, the shared verdict and species if the animal boxes agree, else real with no species (a rat and a bird: each box keeps its own); then the viewer moves on. Boxes are outlined by their tag (green real, grey false, amber wrong species, red none yet; dashed when the tag comes from the visit's), with a ✓ / ✗ / ≠ mark. A real or wrong-species tag has a what was it? field, suggesting names you've used and groups seen here; a wrong-species tag waits for it instead of moving on. ✓ more precise… is for a log that was right but could say more (animal → doves/pigeons, songbirds → house finch): it tags real and waits for the name too (Enter saves and moves on; next › skips), where plain ✓ real moves straight on. The boxes to tag queue walks the real visits whose still still has an untagged box (narrowed to the camera filter): the ones to revisit before the stills are used as training data.
Counts follow the tags, in Analytics' timeline and in activity_summary (so in Claude's answers
too): a visit tagged false is left out, and one you named counts under that name and its group
(a group or class name as itself, otherwise the group the log files that name under).
On a phone the feeds page keeps its top short: paused cameras share one chip, and the filters fold behind one button that says what they're set to (tap it to change them).
The viewer's own tools sit small and together above the tags: ▶ clip, ☆, hide, boxes and full-res ↗.
Favourites. Separate from tags, ☆ in the viewer stars a visit as a keeper. Starred while its clip plays, it keeps that moment and reopens there (paused); starred on the image, it opens on the still. Starred visits show ★ after their name, and ★ favourites in the filter bar lists them all, from any time (the window is ignored) and hidden or not. Nothing cleans up visit media today; any future cleanup must leave favourites alone.
A visit opens on its full-resolution still, even when the grid loaded before the visit ended; only a visit with no frame kept shows its thumbnail (scaled up, without boxes). In the viewer, scroll or pinch to zoom, and drag to pan once zoomed; on a touch screen a double tap zooms in at that spot, or back out. ⤢ fills the window with the viewer, tags and all; a phone always shows it that way. full-res ↗ opens the still in its own tab, filling the window, with the same zoom and pan (a double click or tap zooms in, or fits again; boxes turns the detector's boxes off; jpeg ↗ is the file itself; tag › opens the visit here).
Exclusion zones. In the zones editor (/zones), the exclude tool marks spots that only
ever produce junk: a feeder roof's underside, a patch of hedge. Each zone has a mode (click it to
switch):
- Ignore: anything boxed here (box centre in the zone) is dropped as soon as MegaDetector finds it, before SpeciesNet classifies it, so it never opens or extends a visit and can't keep the camera analysing. For spots no animal can be. (A phantom "bird" on a feeder roof's underside once held a visit open all night: 72% of that camera's analysed frames in a day.) Dropped boxes leave no visit to show again, but the monitor log keeps a periodic tally ("exclusion zones dropped N on …"); to undo, switch the zone to hide unrecognized or delete it.
- Hide unrecognized: a visit SpeciesNet couldn't identify (a bare "animal"), with all its boxes
in the zone, is hidden when it ends (reason
zone): kept, and shown again with one click. An identified animal is never hidden. It can only act at the end: whether a visit is identified is only settled then. It doesn't reduce load.
An exclusion zone can also have a box size range (optional, per zone; min_size / max_size
in zones.json): only boxes whose larger side, as a fraction of the frame, is at least the "from"
and below the "below" count as in the zone. Size separates a fixed object from an animal at the
same spot, which position alone can't:
- from (big fixed object, small animals): on backyard, 99.9% of the ~70k boxes centred on a feeder roof were 0.052 or larger (median 0.075), while birds crossing there boxed at 0.015–0.03. From 0.04 lets a polygon be generous without dropping those birds.
- below (small fixed object, bigger animals): on west_alley, a bolt-loop swinging on a squirrel shield boxed at 0.024–0.032, while a rat sitting on it boxed at 0.05. A small zone over the bolt's swing with below 0.04 drops the bolt and keeps the rat.
To size a zone for something that moves (a swinging part), draw it over the whole spread of its boxes, not one frame's: the labels map shows where past boxes landed. Zones without a size range behave as before.
Any visit can be shown or hidden by hand from the viewer (show / hide), to undo a wrong call
by the gate or a zone. Zones apply to new visits. In zones.json the modes are ignore and
hide_unrecognized (the earlier all / unnamed are still read).
Alert areas. alert_species: [coyote, owls, …] under site: texts SMS_ALERT_TO the first
time one of those is seen on any camera, then stays quiet for 5 minutes per species. That only
works when the label is right, and labels on small or partly hidden animals often aren't: a cat on
a fence can be logged as a bird, a rodent, an animal or even a person. An alert area goes by
place and size instead. Draw one with the alert area tool in the zones editor, and anything
boxed with its centre inside, at least min size of the frame across and not labelled one of its
skip labels or groups, sends a text, whatever its label: cat on fence: bird at west_alley, 9:02am. The text starts with the area's name and carries the label, so you can judge it.
Each area then stays quiet for 5 minutes. New areas start at min size 0.06 (most songbirds box
smaller; a cat on a fence boxed 0.04–0.21) and skip squirrel, rodent, which sit in the same places
at the same sizes as a cat but are usually labelled right. Areas are opt-in, per camera, and need
the same mail settings and SMS_ALERT_TO. In zones.json they're under alert:
{"rect" | "poly", "name", "min_size", "skip": [...]}.
SMS_ALERT_TO can be a plain email address, and that's the reliable choice: several carriers have
shut their email-to-text gateways (AT&T's txt.att.net no longer delivers). An alert email has the
alert as its subject, shows a close-up of the box that set it off at full resolution, and attaches
the whole frame with the box drawn on it. Set WEB_URL (the dashboard's address as your phone
reaches it, e.g. http://studio.local:8080) in the credentials file and it also links to the visit
(/?visit=<number>), which opens it in the dashboard ready to tag.
Drawing zones. The editor zooms (wheel, trackpad pinch, or + / − / fit) and pans
(Space-drag, middle-drag or ✋), on a full-resolution snapshot. Motion ROI, motion mask,
exclusion zones and alert areas can be polygons as well as rectangles: pick polygon, click points, close
with Enter, a double-click or a click on the first point; drag a point to move it, Alt-click
to delete one. The analysis ROI is always a rectangle (the detector needs one). Polygons cost no
throughput: motion zones are rasterised into a pixel mask once at startup (per frame it's one
multiply, whatever the shape), and exclusion zones are checked once per visit. In zones.json a
polygon is {"poly": [[x, y], ...]} beside the [x, y, w, h] rectangles.
Changing the analysis ROI. A visit's boxes are stored as fractions of the analysis ROI in
effect when it was analysed, and each visit records that ROI (visits.roi). So after the ROI
changes, older visits still show their boxes where they were (viewer, labels map, tag model).
Visits from before the ROI was recorded fall back to the camera's ROI now. Once it's known which
ROI was in effect when (draw a few stored boxes on their stills under each), record it on them:
python -m camera_mcp.store backfill-roi --camera west_alley --roi 0.08,0.13,0.48,0.43 --until 2026-11-04T20:38:26Z
--since / --until are UTC; no --roi means the whole frame. It only fills visits that have no
ROI recorded, never overwrites one.
Moved-camera check. Every 5 minutes, when nothing is moving, the monitor compares each
camera's view with a stored reference (one for day, one for infrared night, in data/views/).
If three checks in a row don't match, it logs a warning and sets view_moved_since in the
camera's status: a raccoon may have tilted it. After re-aiming a camera on purpose, forget its
old reference:
.venv/bin/camera-monitor --reset-view east_fence
Preguntas sobre camera MCP
¿Qué herramientas proporciona camera MCP?
Expone 1 herramienta, incluidas list_cameras.
¿Qué pueden hacer las herramientas de MCP?
Las herramientas te permiten listar cámaras, tomar capturas de pantalla, encontrar visitas de animales dentro de rangos de tiempo o taxa específicos, obtener resúmenes de actividad y verificar el estado del monitor.
¿Identifica especies específicas?
Sí, utiliza SpeciesNet para nombrar a los animales, limitado a las especies que se encuentran en tu área especificada.