Bridgit API

Take a photo

An agent can take a photo with the front or back camera. Android's green dot and a Bridgit notification show every time.

• Back or front camera
• Zoom: 0.5× (ultra-wide) up to the phone's maximum (telephoto)
• Flash on or off
• Size: sized for AI (1568 px), thumbnail (512 px), large (2560 px) or original
• Or an exact width or long edge, and JPEG quality
• Resolution supported by the phone, camera pipeline and response limits
• As JSON with the picture inside, or as plain JPEG bytes
• Never saved to your gallery

From your laptop: "Is the stove off?" Your agent takes a photo of the kitchen and tells you.

POST /v1/camera/snapshot {"camera":"rear","zoom":2}

Record a short clip

An agent can record 1 to 120 seconds of audio and get it back as a file.

• Length: 1 to 120 whole seconds
• Comes back as an M4A (AAC) audio file, 44.1 kHz mono
• Not kept on the phone

"Listen to the next 30 seconds of the meeting and give me the action items."

POST /v1/mic/record {"duration_seconds":5}

Your latest screenshot or photo

One call returns the newest screenshot or camera photo, sized for an AI to read. It can also wait for the next one you take.

• Which: any image, the newest screenshot, or the newest camera photo
• Wait up to 55 seconds for the next one you take
• Size: sized for AI (default), thumbnail, large or original
• As JSON or as plain JPEG bytes, with when it was taken and the album

You screenshot an error on your phone; your agent already has it and explains the fix.

GET /v1/image/last?type=screenshot

Find photos in your gallery

Agents can list recent images and fetch one at the size they need. Gallery retrieval does not edit or delete your original gallery media.

• List the newest 1 to 100 images (no pictures, just details)
• Fetch one image at any of the sizes
• Works with full access or only the photos you picked

"Find the receipt photo from this morning and add it to my expenses."

GET /v1/gallery/recent?limit=10

GET /v1/gallery/photo/87567?size=agent

Questions and notices to you

Agents send you a notice, or a question with buttons, choices or a text reply, and receive your answer through the API or callback when the connection is available. They can also read it aloud.

• A notice (one-way) or a question
• Buttons (up to 6), a text reply, a choice list, or a link
• An image with the message
• Normal or urgent; read aloud if you allow it
• Expires after 1 minute to 7 days
• The agent waits for your answer or gets a callback

"Pay this invoice? $11.99" [Pay] [Not now]. Your reply is returned to the agent. Any payment requires an authorized payment tool and confirmation of the actual result.

POST /v1/interactions {"kind":"question","title":"Pay this invoice?","controls":[{"type":"button","id":"yes","label":"Pay","style":"primary"},{"type":"button","id":"no","label":"Not now"}]}

GET /v1/interactions?status=answered

GET /v1/interactions/12?wait=55

DELETE /v1/interactions/12

GET /v1/events?since=0&wait=55

Rich HTML cards

Instead of plain buttons, an agent can send a small HTML page (charts, forms, choices) that runs in an isolated view on the phone without network access.

• A small HTML page with its own layout, charts and forms
• Runs on the phone with no internet access
• Sends one answer back to the agent (up to 16 KB of data)
• Images inside the card only (data: images), up to 256 KB of HTML
• Bridgit's own header shows which agent sent it; you can always dismiss it

Your travel agent shows three flights as cards with prices and a map; you tap one to book.

POST /v1/interactions {"kind":"question","title":"Pay this invoice?","controls":[{"type":"button","id":"yes","label":"Pay","style":"primary"},{"type":"button","id":"no","label":"Not now"}]}

GET /v1/interactions/12?wait=55

Live text panel

An agent streams text into the lower half of the screen and it scrolls as it arrives, like the transcript of an AI session. Pause the scrolling or close it any time.

• Lower half of the screen, scrolls as text arrives
• Pause scrolling to read, then resume
• The agent marks it finished; you close it whenever you like
• Keeps the last 200 KB of text

Claude Code is fixing a bug on your laptop; you watch what it's doing on the phone while you make coffee.

POST /v1/interactions {"kind":"question","title":"Pay this invoice?","controls":[{"type":"button","id":"yes","label":"Pay","style":"primary"},{"type":"button","id":"no","label":"Not now"}]}

POST /v1/text/stream {"interaction_id":"42","append":"Claude: running the tests now...\n"}

DELETE /v1/interactions/12

Send a screenshot

An agent sees what's on your screen right now, with no prompt each time; a notification shows every capture.

• A private copy: nothing is saved and there's no flash
• Turn on Bridgit's screenshot service once (Access), then Screenshots in Features
• Same size choices as photos
• At most one every 2 seconds and 60 an hour per agent

"What does this error on my phone mean?" The agent takes a screenshot and reads it.

POST /v1/screen/capture?size=agent

Open a link or app

An agent can suggest a link or app; it opens when you tap it.

• Web links, maps, phone numbers (dialler only), email
• Or open an app by name
• Only when you tap it

"Here's the boarding pass" [Open in Wallet].

POST /v1/interactions {"kind":"question","title":"Pay this invoice?","controls":[{"type":"button","id":"yes","label":"Pay","style":"primary"},{"type":"button","id":"no","label":"Not now"}]}

Phone status and context

Battery, network, ringer, Do Not Disturb and screen/lock status. An unlocked interactive screen is reported as in use.

• Battery level and charging
• Screen on/off and lock state
• Ringer, Do Not Disturb and media volume
• Wi-Fi/mobile transport and metered status
• Context: in use when screen is interactive and unlocked; otherwise unknown

Before calling you, an agent checks whether your phone is in use or on silent.

GET /v1/phone/status

GET /v1/phone/context

Location details

Where the phone is, on request only: latitude and longitude (x, y), altitude (z), accuracy, speed and heading, and how old the fix is. Address lookup if you turn it on. No tracking unless you choose it.

• Latitude and longitude (x, y)
• Altitude (z) in metres
• Accuracy in metres, speed and heading
• A fresh fix (up to 15 s) or the last known one at once
• High accuracy or battery-friendly
• Street address, if you turn on address lookup

"How long until I'm home?" The agent checks where you are and the traffic.

GET /v1/location

Flash the light

Turn the torch on for up to ten minutes, with automatic off and an owner Stop button.

• On for 1 second to 10 minutes; always turns itself off
• Turn off now
• Ask whether it's on

Your door-camera agent flashes the light so you know someone is at the door.

GET /v1/flashlight

POST /v1/flashlight

Phone controls

Do Not Disturb, ringer, volume and brightness, put back as they were afterwards.

• DND: off, priority, alarms or none; none can mute alarms
• Ringer: normal, vibrate or silent; media/ring/alarm volumes
• Brightness level or automatic brightness
• Optional 1–1440 minute restoration; owner Undo
• Later manual changes are preserved

Bedtime routine: your agent turns on Do Not Disturb until 7 am.

POST /v1/phone/dnd

POST /v1/phone/sound

POST /v1/phone/brightness

POST /v1/phone/undo

City tour

A guided walk: you pick the kind of tour, and every minute your agent gets where you are, which way you face and the photos you just took. It tells you about what's around you, and you can ask it questions out loud.

• History, architecture, environment, food, art, or anything interesting
• Updates every 30 s, 1, 2 or 5 minutes
• Location, heading (compass) and new photos each time
• Ask out loud or type a question
• Pause, skip, end; ends by itself after 4 hours

Pick "Architecture" and walk; when you photograph a church, your agent tells you who built it and why it looks that way.

GET /v1/tour

DELETE /v1/tour

GET /v1/events?since=0&wait=55

POST /v1/speaker/say {"text":"On your left is the old town hall, built in 1360.","tour_id":"t123"}

POST /v1/speaker/play (Content-Type: audio/mpeg, body = the MP3)

GET /v1/gallery/photo/87567?size=agent

Find my phone

Ring and vibrate for up to a minute, respecting your sound and Do Not Disturb settings.

• Ring and vibrate for 1–60 seconds
• Respects Android sound and Do Not Disturb settings
• Stop from Device tools or the ongoing notification

Phone lost in the sofa: ask any agent to ring it.

POST /v1/phone/ring

DELETE /v1/phone/ring

Sensors

Light, pressure, steps, motion and orientation, one reading each.

• Light, pressure and proximity where available
• Steps since reboot (Physical activity grant)
• Fresh rotation-vector orientation and accelerometer values
• Unavailable or stale sensors are reported explicitly

An agent notices it's dark and you're still, and holds non-urgent messages.

GET /v1/sensors

Videos and downloads

Find recent videos, take a frame from one, and pick a file from Downloads for an agent.

• List accessible videos
• Retrieve a video up to 40 MB, or a still frame
• Choose individual Downloads files on the phone for one agent

"Get the frame at 0:12 from the video I just took."

GET /v1/gallery/videos/recent

GET /v1/gallery/video/{id}

GET /v1/downloads

Files to your phone

Agents send PDFs, images and other files; they land in the Inbox and can be saved to Downloads.

• 25 MB per file; 100 files / 100 MB total
• Private storage, retained 30 days
• Owner opens, exports to a selected location, or deletes
• No automatic execution or automatic Downloads save

Your accounting agent sends this month's invoice PDF to your phone.

POST /v1/files

GET /v1/files

Share to an agent

Share a photo, link or text from any app straight to Grok or Claude.

• Selected content files, text and links from Android sharing
• Choose which permitted agent receives each share
• Agent reads its own transfer history
• Up to ten selected files, each limited to 25 MB

Share a restaurant link to your agent: "Book this for Friday."

GET /v1/shares

GET /v1/shares/{id}

Short video clips

1 to 30 seconds of 720p or 1080p MP4. Audio is off by default and needs a microphone grant.

• 1 to 30 seconds
• Back or front camera
• With or without sound
• 720p or 1080p

"Record the dishwasher noise so I can describe it to the repair shop."

POST /v1/camera/video

Read text on the phone

Bundled Latin-script text recognition runs on the phone; only the result leaves it.

• Bundled Latin-script recognition
• From an accessible gallery image or supplied image bytes
• Returns recognized text; no cloud recognition

Read the serial number from a photo without uploading the picture.

POST /v1/ocr

Transcribe clips on the phone

Request transcribe=true on a microphone clip. Requires Android on-device recognition and an installed offline language; no cloud fallback.

• Transcript with the audio, or the transcript only

Dictate a note for 20 seconds; the agent gets the words.

POST /v1/mic/record {"duration_seconds":5}

Screen sharing as a fallback

If you'd rather not turn on Accessibility, approve screen access once, for 1 hour or for 16 hours.

• Allow once, for 1 hour or for 16 hours, or deny
• Stop any time from the status bar

Help your agent debug an app for the next hour.

POST /v1/screen/grant

GET /v1/screen/grants

POST /v1/screen/share/capture

DELETE /v1/screen/grant

App usage

Read foreground-time summaries after granting Android Usage access. Android buckets can extend past the requested interval.

• Foreground-time summaries for up to seven days
• Requires Android Usage access
• Android buckets may extend outside the requested interval

Weekly review: "You spent 6 hours in email this week."

GET /v1/usage

Health Connect

Read permitted steps, sleep and heart-rate source records. No health writes; maximum 100 records per query.

• Read permitted steps, sleep sessions or heart-rate samples
• No writes, background reads, workouts or retained health records
• At most 100 source records and seven days per request
• Source records may overlap; no deduplicated total claim

Your coach agent adjusts tomorrow's plan to last night's sleep.

GET /v1/health/records

Read notifications

Read up to 100 active notification texts received since the listener connected. Explicit notification access required; memory only.

• Explicit feature switch and Android notification access
• At most 100 active notifications since listener connection
• RAM only; cleared when disabled, paused or disconnected
• No notification-history archive or long-poll in this build

"Summarise everything that pinged me during the meeting."

GET /v1/notifications

Copy to clipboard

An agent puts text on your clipboard after you approve it.

• Up to 10000 characters of text
• Every copy requires your foreground Copy tap
• Five-minute approval expiry
• Text is marked sensitive to suppress clipboard previews

Your agent copies the 2-factor backup code so you can paste it.

POST /v1/clipboard

GET /v1/clipboard/{id}

Play an audio file

An agent sends a sound, a voice message or text to read out, and the phone plays it on its speaker. Newer audio from the same agent replaces what's waiting, so it keeps up with you.

• An MP3, M4A, OGG, WAV or FLAC file (up to 25 MB), or a public https link
• Or text read aloud with the phone's own voice: language, speed and pitch
• Replace what's waiting, stop the current one, or queue
• Never louder than your volume ceiling (default 70 %)
• Silent when your phone is on silent, vibrate, Do Not Disturb or a call
• Stop clears everything

Your reminder agent plays "Time to leave for the airport" in its own voice.

POST /v1/speaker/play (Content-Type: audio/mpeg, body = the MP3)

POST /v1/speaker/say {"text":"On your left is the old town hall, built in 1360.","tour_id":"t123"}

GET /v1/speaker

DELETE /v1/speaker

Listen for a word on the phone

On-device word spotting while you allow it, sending only what follows the word. Needs a change to the "no always-on listening" rule.

Say "Bridgit, note" and the next sentence goes to your notes agent.

Not built yet. Spec PRD-09 in docs/bridgit-spec-v2.md describes the planned calls.

Messages, contacts, calendar, calls

Read and, with your approval each time, send.

"Text Sam I'm 10 minutes late" — you approve on the phone.

Not built yet. Spec 8.22, Q1 in docs/bridgit-spec-v2.md describes the planned calls.

Location history and places

Optional history kept on the phone, and "tell me when I get home".

Grok reminds you to water the plants when you arrive home.

Not built yet. Spec CAP-20a, Q14 in docs/bridgit-spec-v2.md describes the planned calls.

Add an agent in a minute

Bridgit makes a setup prompt with the agent's own key, the address and the rules. Use it with a compatible agent whose HTTP or MCP tools can reach the phone’s Tailscale network.

You add Claude: name it, copy the prompt, paste it into Claude. The screen turns to "Claude connected" as soon as it calls.

GET /v1/health

POST /v1/pair {"code":"482913"}

GET /v1/tools/bridget

One key per agent

Every agent has its own key. Rotate it or remove the agent in one tap; the old key stops at once and the agent is told why.

You pasted Grok's prompt into the wrong chat. Rotate Grok's key; the leaked one is dead in a second.

GET /v1/health

Everything is logged

Every request, switch change and agent note is in Activity on the phone. Agents can read and follow the log too, so they know what they did.

• Filter by agent, type, endpoint, result, level, time or text
• Follow it live
• Stats per agent and endpoint
• Export as JSONL or CSV
• "All entries" or "only their own" for agents

Codex checks yesterday's log to see why a photo failed; you scroll Activity to see who did what.

GET /v1/log?agent=me&limit=20

GET /v1/log/48213

GET /v1/log/stats?since=2026-10-27T00:00:00%2B01:00

GET /v1/log/export?format=csv&agent=me

Agent notes

Agents can write short notes into the log.

Your fridge-check agent writes "milk low" so you see it in Activity.

POST /v1/log {"message":"checked fridge photo, milk low"}

The API explains itself

Agents read a live guide and an OpenAPI file that list what's switched on, every call and every error.

A brand-new agent reads the guide and learns everything without a manual.

GET /v1/guide

GET /v1/openapi.json

GET /v1/capabilities

Agents learn what's new

Every reply says which API revision Bridgit speaks. After an update, agents are told there's something new until they read the list of changes, so they pick up new features by themselves.

Bridgit gains camera zoom; the next time Grok calls, it reads what's new and starts using zoom.

GET /v1/changes?since=2

Report a bug or ask for a feature

From any screen: the last screen is attached, plus screenshots and the recent log. Developer agents read reports and mark them fixed; you confirm or say it's still broken.

• Last screen, up to 5 screenshots and the recent log attached
• "Don't auto-fix" keeps it for a person
• Statuses: new, seen, in progress, fixed, won't fix, duplicate
• You confirm a fix or say it's still broken

The app misbehaves; you report it in 10 seconds. Claude picks it up, fixes it and marks it fixed.

GET /v1/feedback?status=new

GET /v1/feedback/7

POST /v1/feedback/7/status {"status":"in_progress","note":"Looking at it"}

POST /v1/feedback {"kind":"bug","description":"image/last returned 500 twice"}

Check for updates

Bridgit finds a new version, checks it really is Bridgit, and lets Android install it after you confirm.

A fix ships; your phone offers the update the next time you open Bridgit.

You stay in control

Every feature starts off and has its own switch. Pause listening in one tap. Access shows each Android permission with one button to the right page.

Going into a meeting: turn off Camera and Microphone from Home in two taps.

GET /v1/capabilities

Only over Tailscale

Bridgit answers only on your private Tailscale network, never the open internet.

Your agents on your Mac and your server can reach the phone; nobody else can.

Use Bridgit over MCP

Agents that speak MCP (Claude Code, Codex and others) add Bridgit as a tool server with their own key, and every feature appears as a tool.

• Same key, same switches, same log as the normal API
• Every endpoint the agent may use becomes a tool, with its options
• Photos and screenshots come back as images

Add Bridgit to Claude Code once; from then on "take a photo of my desk" just works.

POST /v1/mcp {"jsonrpc":"2.0","id":1,"method":"tools/list"}

Inbox

One place for every question, notice, file and share between you and your agents.

See what Grok asked you this morning and what you answered.

GET /v1/interactions?status=answered

GET /v1/files

GET /v1/shares

Events and callbacks

Agents get told the moment something happens: a question answered, a new screenshot, a switch changed.

Your agent waits for your answer without asking again and again.

GET /v1/events?since=0&wait=55

POST /v1/callback/test

Quick Settings tile

Pause or resume Bridgit from the notification shade.

Pull down, tap Bridgit, and agents can't reach the phone until you tap again.

Per-agent permissions

Choose what each agent may use, for example photos for Claude but not for Grok.

Let your shopping agent see the gallery but never the camera.

GET /v1/capabilities

Reach Bridgit without Tailscale (relay)

Agents that can't join Tailscale, like hosted assistants, leave jobs in an encrypted mailbox; a push wakes the phone.

Muse, which runs on Meta's servers, takes a photo for you.

Not built yet. Spec 4.4, Q10 in docs/bridgit-spec-v2.md describes the planned calls.

A. Run it from your side (no app change)

Any paired agent can do this today with the calls below.

1. Report your model once: POST /v1/me {"model","harness"}.
2. Offer the applet with a question card: POST /v1/interactions with buttons, e.g. Start / Not now. Wait for the answer with GET /v1/interactions/{id}?wait=55.
3. Keep one loop on GET /v1/events?since=<last id>&wait=55 (60 s timeout). If your harness only acts during a turn, run the wait as a background command so each event wakes you.
4. Read what the applet needs on a timer you keep: GET /v1/location (with compass), POST /v1/camera/snapshot, GET /v1/image/last?type=photo, POST /v1/mic/record. Check GET /v1/capabilities first and ask the owner to switch on what is off.
5. Talk back with POST /v1/speaker/say {"text"} (2-4 short spoken sentences; newer narration replaces what is queued) or a question card when you need a choice.
6. Stop when the owner says so (a Stop button on your card) or after a time limit you announced. Only capture while the applet runs; every capture is shown and logged.

B. Add it to the Bridgit app (a coding agent with the repository)

Copy City tour. Each step names the City tour file to copy from; keep its rules.

1. Write the spec first: a section next to 8.25 City tour in docs/bridgit-spec-v2.md with numbered rules like TOUR-01..07 (what it needs switched on, what is sent to whom, limits).
2. Session engine: a manager like TourManager in tour/Tour.kt. One session at a time, one agent, sends <applet>.started (with a brief string telling the agent what to do), <applet>.tick, owner actions, and <applet>.ended through app.interactions.event(agentId, type, data). Create it in BridgeApp.
3. Agent API: a Capability module like TourModule (register it in capability/Capability.kt), with an EndpointSpec for every call in api/Endpoints.kt and in Endpoints.all. Only the session's own agent may read it.
4. Screen: a fragment like ui/tour/TourFragment.kt (setup: pick an agent and options, list what is missing; running: big buttons, what was said, End). Add it to res/navigation/nav_main.xml.
5. Card: add it to res/layout/fragment_applets.xml and AppletsFragment, with a running chip.
6. Tell agents: add an ApiChanges revision, a section in the live guide (api/SelfDescription.kt), and a "What can Bridgit do?" card in ui/more/WhatCanData.kt.
7. Use the automatic APK version allocator; every APK build must reserve a fresh version name and code. Do not hardcode or lower the version counter, run ./gradlew testDebugUnitTest (ApiDocsTest checks every endpoint is documented), install, and try it on the phone with a real agent.

Rules every applet keeps: data goes only to the session's agent and only while it runs; the owner can always end it; nothing is kept afterwards except log lines.