claude-gateway

Reply on Telegram. Pass chat_id from the inbound message. Optionally pass reply_to (message_id) for threading, and files (absolute paths) to attach images or documents.

Add an emoji reaction to a Telegram message. Telegram only accepts a fixed whitelist (👍 👎 ❤ 🔥 👀 🎉 etc) — non-whitelisted emoji will be rejected.

Download a file attachment from a Telegram message to the local inbox. Use when the inbound <channel> meta shows attachment_file_id. Returns the local file path ready to Read. Telegram caps bot downloads at 20MB.

Edit a message the bot previously sent. Useful for interim progress updates. Edits don't trigger push notifications — send a new reply when a long task completes so the user's device pings.

Create a new Claude Gateway agent with a workspace and channel configuration. The gateway hot-adds the agent without requiring a restart.

Update an existing agent: add a new channel, remove a channel, or update a workspace file.

Browse the community app registry. Omit name to list all apps; provide name to get versions for a specific app.

Install an app from the registry or a GitHub URL. Returns a jobId to poll with poll_install_job.

Poll the status of an app install job. Returns status (pending/running/completed/failed), logs, and result on completion.

Preview an install source WITHOUT installing: fetch and parse the app.yaml and return the required secrets (secretKeys, must be provided) and auto-generated secrets (generatedKeys, filled by the gateway), plus name, version, ports, and warnings. Also returns secretDefaults — defaults for prompt-with…

List all installed apps with their status and proxy URLs.

Get detailed status and version info for an installed app.

Update an installed app. Registry apps update to their latest published version; GitHub-URL (custom) apps update to their repo default-branch HEAD. Returns a jobId.

Uninstall an app — stops containers, removes images, deletes files.

Start, stop, or restart an installed app.

Reclaim leaked Docker build cache and dangling images left behind by app install/update. mode "report" (default) returns a read-only reclaim report: reclaimable build cache, dangling image count, and orphan volume names. mode "prune" executes ONLY the safe reclaim (build cache older than the configu…

Back up an installed app — a permission-safe snapshot of its Docker named volumes, bind-mount data directories under the app dir, and config (.env/app.yaml) into a single archive. The app is briefly stopped for a consistent snapshot and restarted afterward. Returns a jobId; poll with poll_install_jo…

Restore an app's Docker named volumes, bind-mount data directories under the app dir, and config from a prior backup, then start it on the restored data. Restoring a backup from a different app version is allowed but may warn about schema/migration mismatch. Returns a jobId; poll with poll_install_j…

List an app's backups, newest first: each entry has id, createdAt, sizeBytes, and appVersion. Use the id with restore_app.

Create or resume GetPod Cloud Browser, NOT the user Chrome or Remote Browser connector. Use only when Cloud Browser is selected, or no Remote Browser is connected for a generic browser request. Returns session status.

Close a browser session, killing the process and removing session data.

Get the WebSocket stream URL for an active session (for frontend live view).

Navigate to a URL. If tab_id is provided, navigates that specific tab; otherwise navigates the active tab.

Return the accessibility tree of the current page.

Click an element by accessibility ref or CSS selector.

Fill an input element with a value.

Type text into the currently focused element.

Evaluate JavaScript in the browser and return the result.

Scroll the page at (x, y) by (deltaX, deltaY).

Wait for a condition: element selector, networkidle, or URL pattern.

Get text content of an element matching the selector.

Open a new browser tab, optionally navigating to a URL. Returns tab_id.

Close a browser tab by tab_id.

List all open browser tabs.

Capture the current viewport as JPEG. Returns the absolute file path of the saved image (ready to attach to Telegram).

Navigate a specific tab (by tab_id) to a URL.

List scheduled cron jobs for this agent

Create a new cron job

Delete a cron job by ID

Update an existing cron job. Provide job_id plus only the fields to change.

Run a cron job immediately

Get run history for a cron job

Send a message to a Discord channel, thread, or DM. Pass channel_id from the inbound message. Optionally pass reply_to (message_id) and files (absolute paths).

Add an emoji reaction to a Discord message.

Edit a message the bot previously sent.

Download a file from a Discord CDN URL to the local inbox.

Create a public thread in a Discord channel.

Search your long-term memory archive (past notes under memory/, plus MEMORY.md/USER.md) by keyword and get the most relevant snippets with their file + line range. Use this to recall details that are NOT in the memory currently loaded into your context — the archive holds far more than fits in the p…

Read an exact line range from one of your memory files (a path returned by memory_search, e.g. "memory/foo.md", or "MEMORY.md"/"USER.md"). Use after memory_search to pull the full context around a snippet. Read-only.

Create a NEW note in the shared, cross-agent knowledge base RIGHT NOW, instead of waiting for the nightly automatic promotion. This is an explicit, agent-initiated write (works even when shared-KB mode is "propose"). Notes live in a shared namespace across every agent — pick any "name" you like, the…

Read the full current content of one shared-KB note by its exact "name" (see memory_shared_create). Use this before memory_shared_update so your edit is based on what is actually there — memory_search only returns short snippets, not the full note.

Delete one shared-KB note by its exact "name" (see memory_shared_create). Any agent can delete any note in the shared vault — there is no per-agent ownership. This cannot delete notes the nightly dreaming pipeline promoted (different naming scheme).

Create a new skill from content. The skill will be available immediately via hot-reload.

Delete a skill by name.

Install a skill from a GitHub URL or raw URL pointing to a SKILL.md file.

Read the skill self-improvement effectiveness rollup for this agent: adoption funnel (auto-skill stickiness), cost-to-complete deltas, recovery-rate trend, enabled/disabled cohort, and the net-token ledger (tokens saved by reuse minus tokens spent reviewing). Read-only.

Discover all capabilities this agent can offer: installed/shared/native skills, gateway MCP tools and enabled connector/plugin MCP tools, including tools executed only by workers. Read metadata only; never executes a tool or grants permissions. Use query to find relevant tools before assigning work.…

Prepare executable work, incomplete materials, or an amendment to an existing task. Do not call this for greetings, introductions, casual conversation or questions you can answer directly; answer those normally once. When semantic intake is enabled, acknowledge before work mutations. Incomplete mate…

Durably queue a task for an independent worker. Returns after commit without waiting for worker startup or completion.

Read the task index in this conversation without waiting. Pass task_id to retrieve the complete stored result and evidence before discussing an earlier task. An index entry is not a partial worker report.

Request cancellation, including needs_reconciliation cleanup. cancel_requested is not proof of termination. Existing files are retained. Set replaced_by_task_id only for a known replacement task, never infer from similar titles.

Accept a revised task instruction. applied_revision changes only when the worker accepts it.

Manage pending questions without answering or resuming work. Ask stages one natural separate message after your reply; discuss marks ongoing consultation; defer/mute/resume persist the user reminder preference.

Answer the matching pending question on a task.

Append an explicitly requested memory note for a channel task. API tasks cannot write memory. Files are restricted to MEMORY.md, USER.md, or memory/<name>.md. Retains existing content and deduplicates retries.

Stage a finished image or document for the user. The orchestration attaches it to the next completed agent response after this task succeeds, using the original destination. This tool does not send to a channel. Call once per output file. For an image returned by an MCP tool (such as a remote browse…

Report factual progress internally. At phase changes include a checkpoint with phase, evidenceVersion, checks, findings and nextAction. This does not send a user message.

Ask the user for information required by your task, then end this turn immediately.

Reply on Telegram. Pass chat_id from the inbound message. Optionally pass reply_to (message_id) for threading, and files (absolute paths) to attach images or documents.

Add an emoji reaction to a Telegram message. Telegram only accepts a fixed whitelist (👍 👎 ❤ 🔥 👀 🎉 etc) — non-whitelisted emoji will be rejected.

Download a file attachment from a Telegram message to the local inbox. Use when the inbound <channel> meta shows attachment_file_id. Returns the local file path ready to Read. Telegram caps bot downloads at 20MB.

Edit a message the bot previously sent. Useful for interim progress updates. Edits don't trigger push notifications — send a new reply when a long task completes so the user's device pings.