Bird vs Sinch (2026)

An agent signs in to Bird through browser OAuth and then acts with a scoped, revocable grant. Through the hosted MCP server it sends email, SMS, WhatsApp messages and one time passcodes, orders phone numbers, adds sending domains and senders, and manages templates, contacts, audiences and preferences.

Grants cover only the overlap of what the client asked for, what the user approved and what the user holds, and owner and platform admin scopes can never be delegated. In the other direction, webhook destination connectors hand each inbound message to a Claude Managed Agents session or a Grok routine. Connections to the customer's own business systems are the customer's agent's to make.

Sourcebird.com/docs/ai/mcp-serverread 2026-10-11

Action blocks connect a bot to outside systems in three ways. An API step sends requests to any HTTPS endpoint with any of the five standard HTTP methods, authenticates with basic auth or a bearer token, passes session variables in the query or body, and can route the conversation to the next block named in the response. Code actions run JavaScript that calls external services and writes the results back to the session.

Native app steps create, find and update Zendesk tickets and users, create Salesforce contacts, accounts and cases or run SOQL queries, and connect Freshdesk, Topdesk, Gmail and Intercom. In a banking bot, an API step can post a money transfer's amount, destination and account type to the customer's back end in the middle of a conversation.

Sourcedocs.chatlayer.ai/integrateandcode/custom-back-end-integrations/configure-your-api-integrationread 2026-10-11

Events move to the agent and its output moves back out through Bird. Each inbound WhatsApp, SMS, Apple Messages or mailbox message is delivered as event JSON, failed deliveries are retried, and a webhook can be paused or deleted.

The sequence of steps an agent takes, its tools and its decisions run in the customer's agent platform, and for voice the customer's application owns speech, reasoning and business tools while Bird carries the call.

There is no workflow builder or multistep runtime for agents inside Bird, so a reminder sequence or an escalation path is something the customer's agent or its own scheduler runs.

Sourcebird.com/docs/ai/route-messages-to-an-agentread 2026-10-11

Bots run as flows on a visual canvas, where each block sends a message, collects input, checks a condition or runs an action, and go to connections move the conversation between blocks and flows.

Condition blocks branch on variables, an API response can name the next block, events start a flow when a variable changes or the user goes quiet, and delay and pause steps hold the conversation. Flows can be reused across a bot and published one at a time. The intent engine picks the flow and the designer sets every step inside it, so the sequence is fixed when the bot is built. Running two bots in one conversation takes the web widget or a custom proxy over the webhook channel.

Sourcedocs.chatlayer.ai/navigation/bot-builder/flowsread 2026-10-11

Inbound WhatsApp, SMS, Apple Messages and agent mailbox email each fire an event, whatsapp.received, sms.received, amb.received and email_mailbox.message_received, that a webhook delivers to the agent and that starts a new agent run. MCP Events let an MCP client subscribe to events such as new mail and receive each as a signed webhook. Voice calls reach a customer built voice application over SIP, and agent mailboxes give an agent an address of its own.

Sourcebird.com/docs/ai/route-messages-to-an-agentread 2026-10-11

Conversations arrive through the Sinch Conversation API on WhatsApp, RCS, SMS, MMS, Messenger, Instagram, Viber, Telegram and LINE, through the Web V2 widget, through a webhook channel for any custom front end, and through phone lines for voicebots.

Inside a conversation, events start a flow when a variable changes, when the user stays silent for anywhere from 1 to 1,440 minutes, or when a WhatsApp Flow is completed.

Outside systems can also push into a live conversation through the REST API, sending messages, setting session data or triggering a specific block for a given user, so a back end event can move a conversation forward.

Sourcedocs.chatlayer.ai/channels/all-channelsread 2026-10-11

No knowledge base for an agent to search sits inside Bird. Its skills are instruction packs for the agent itself, one routing requests to the right CLI command group and one auditing a sending domain's authentication and mail records, and the knowledge base Bird runs is customer help for account setup, deliverability and billing. Grounding an agent in a company's own documents happens in whatever agent platform the customer runs. Templates, contacts and audiences in Bird hold messaging data, which an agent can read through the MCP server but which is not indexed for retrieval.

Sourcebird.com/docs/ai/set-up-your-agentread 2026-10-11

Knowledge base AI keeps a store of the customer's content and generates short answers from it. Content goes in as PDF files, CSV question and answer files, single web pages or entire domains, and tags limit which content a given flow searches.

Each answer comes back with the retrieved passages, the source file or URL and the specific page it came from, so a bot can offer a link to the source. When nothing matches, a no findings path sends the user to another block instead of inventing an answer, and a non generative mode replies only with the exact answers in a CSV file. The feature is free up to 50 questions a month, with higher volumes through sales.

Sourcedocs.chatlayer.ai/navigation/knowledge-base-airead 2026-10-11

Every routed message starts a new session or run, and the agent receives the event payload for that message alone, so any history across messages lives in the customer's agent platform. Message records, contacts and audiences persist in the organization's region as the messaging data model, and an agent can look a contact up through the MCP server. Nothing in Bird stores what an agent learned about a person or hands it back on the next message.

Sourcebird.com/docs/ai/route-messages-to-an-agentread 2026-10-11

Each user's session holds what the bot has learned about them, including variables it collected, API results and the last intent, and one session can span several conversations. Sessions expire after 30 days without activity by default, a period Enterprise customers can change per bot and separately for the draft and live versions.

Through the REST API a team can read a session, write to it, delete it for one conversation or delete all data for one user. Bots use saved variables to recognize returning users and reuse earlier answers, such as a traveler's usual route. Persistence depends on the channel, and the web widget keeps variables for one visit unless users are identified. Variables marked sensitive stay out of conversation history.

Sourcedocs.chatlayer.ai/bot-answers/sessionread 2026-10-11

A person approves each agent's access on a consent screen that flags the client name as unverified, can revoke the grant at once under connected apps, and can pause a webhook to stop deliveries.

The outbound call tool prepares a call that a person reviews and places from the browser, template versions reach recipients only once submitted, and destructive tools carry annotations so the agent host can ask before running them.

Sends of email, SMS and WhatsApp go out when the agent calls the tool, with no Bird approval step in between, and handing a conversation to a person is left to the customer's own systems.

Sourcebird.com/docs/ai/mcp-serverread 2026-10-11

A Send to offload provider step hands a conversation to a person in Sinch Contact Pro, Genesys Cloud, Salesforce Service Cloud, Zendesk Chat, Help Scout or any platform behind an offloading webhook, with the transcript and collected variables passed along. The bot keeps answering until an agent accepts, then pauses until the agent sends the user back.

Designers choose where handover happens, typically on request, after repeated not understood turns or after two unsatisfactory knowledge base answers. Generative features can be switched off per bot, and a split between draft and live versions keeps unpublished changes away from customers. No step lets a person approve an action or a reply before the bot sends it.

Sourcedocs.chatlayer.ai/integrateandcode/human-offloading-live-chatread 2026-10-11

Teams are organized into workspaces with member roles, API keys are scoped per channel with none, read or write access to email, SMS and WhatsApp, and agent access runs on OAuth grants limited to what the user holds and approved. Bird holds ISO/IEC 27001:2022, SOC 2 Type 2 and SOC 3 reports and HIPAA compliance, and its trust center supplies an ISO certificate bundle, statement of applicability and penetration test letters on request. Bird does not mention single sign on for its dashboard.

Sourcebird.com/securityread 2026-10-11

Sign in runs through Sinch ID, and SAML 2.0 single sign on with Azure Active Directory, Okta, OneLogin or Ping Identity comes with the Enterprise pack. Admins see every bot and manage the team, members see open bots plus the restricted bots they are granted, and each restricted bot gives a member one of five permissions, Edit, Read only, NLP only, Bot Builder or Copywriter.

Legacy password accounts lock after 10 failed sign ins. Chatlayer is ISO 27001 certified, and across its products Sinch holds ISO 27001:2022 and a SOC 2 Type II report, with PCI and HIPAA compliance covering specific products. Security reports and 343 answered questionnaire items sit in the Sinch Trust Center behind an access request.

Sourcedocs.chatlayer.ai/support/access-controlread 2026-10-11

Audit logs record changes such as API key updates and member invitations, each webhook attempt shows the agent platform's response status and body or a failure reason, and call records and voice events track phone calls, with every message kept as a record in its region. That shows what reached the agent and what Bird sent. The agent's own reasoning and tool sequence are logged in the platform that runs it.

Sourcebird.com/securityread 2026-10-11

Conversation history keeps a turn by turn transcript of every conversation with the intents triggered along the way, voice recordings where recording is on, and tags and filters by user, channel, campaign and date.

Execution logs record each app integration run with its input and output payloads, status, account and the block that triggered it, plus failed API and code actions with the request, status code and response, tracking events and voice speech to text failures. An activity log on the Enterprise pack shows who accessed or changed a bot, what they did and when, and every published live version is kept for rollback. A conversation flow view, labeled beta, traces a conversation's path through blocks for debugging.

Sourcedocs.chatlayer.ai/navigation/history/execution-logsread 2026-10-11

Each organization picks the us1 or eu1 region at signup, starting from a guess based on location, and the choice is fixed afterward. The workspace, API keys, messages, recipient data and event logs stay in that region and are never copied across, with only sign in and account administration held globally. Keys carry their region in the prefix, and a request to the wrong region is refused with a 421 error.

Sourcebird.com/docs/api/regionsread 2026-10-11

Bots run in one of three regions on Google Cloud, each with its own app URL and sign in endpoint, in St. Ghislain, Belgium for the EU, Ashburn, Virginia for the US and Mumbai, India for Asia. Moving bots between regions goes through support, and fixed ingress and egress IP ranges per region let customers allowlist Chatlayer in their own networks.

Sinch's subprocessor register adds Brazil and India as Chatlayer hosting locations, and the Tables API has a South America endpoint, so some data can sit outside the three bot regions. Generative replies through the OpenAI integration send conversation data to OpenAI services in the United States.

Sourcedocs.chatlayer.ai/support/saas-regions-and-ip-rangesread 2026-10-11

Agent skills and plugins come for Claude Code, Codex, Cursor, OpenCode and Antigravity, including a CLI router skill and an email audit skill, plus AI builder guides and a setup prompt to paste into an existing agent. These equip an agent the customer already runs. Bird offers no ready made agents for support or sales.

Sourcebird.com/docs/ai/set-up-your-agentread 2026-10-11

New bots can start from templates built by Chatlayer's conversation design team, each a working bot for one job.

The set covers a banking assistant that makes payments, checks balances and card limits and reports lost cards, ecommerce returns, an electric bike shop that books test rides and checkups, shared scooter support, a bilingual gym assistant that books and cancels classes, job applications, lead generation, NPS surveys, restaurant and retail bots, a nearest location finder and a knowledge base AI bot.

Most work as soon as they load, while those using an integration need an account connected first. Smaller templates add a single function, such as feedback collection, to a bot that already exists.

Sourcedocs.chatlayer.ai/start-quickly/bot-templatesread 2026-10-11

No model runs inside Bird. The customer's agent platform, whether Claude, Codex, Cursor, Grok or its own voice application, reaches Bird through the CLI, MCP or webhooks, and which model that agent uses is decided entirely outside Bird. The webhook connectors are set up for Claude Managed Agents and Grok, and any other platform receives the same event JSON through a plain webhook.

Sourcebird.com/docs/ai/set-up-your-agentread 2026-10-11

An OpenAI action lets a builder pick GPT-3, ChatGPT or GPT-4 for a generated reply, using Chatlayer's own key for the first 100 messages per bot and the customer's OpenAI key after that.

The other generative features, knowledge base AI, the AI intent booster and expression generation, run on unnamed models Chatlayer picks, with OpenAI and Microsoft Azure as Chatlayer's machine learning subprocessors. Intent recognition runs on Chatlayer's own language engine, trained on each customer's expressions.

OpenAI is the only model provider in the integration catalog, and the model list in the OpenAI action still reflects the GPT-3 era.

Sourcedocs.chatlayer.ai/integrateandcode/app-integrations/openai-gpt-3-chatgpt-and-gpt-4read 2026-10-11

Every public endpoint sits in an API reference generated from an OpenAPI spec, with SDKs for TypeScript, Python and Go and a bird CLI for the shell. The MCP server comes in three forms, a hosted endpoint with the full tool catalog, a dynamic endpoint exposing only search and execute for clients with tool limits, and a local server inside the CLI. All three sign in through OAuth with dynamic client registration and PKCE, and their tools create and change Bird's own objects, such as messages, numbers, domains, senders, templates and contacts.

Sourcebird.com/docs/ai/mcp-serverread 2026-10-11

A public REST API with an OpenAPI 3.0 specification covers messaging, conversations, bots and NLP.

It sends messages through the webhook channel with asynchronous callbacks, lists conversations and messages, reads and deletes session data, deletes a user's data, executes actions in a live conversation, lists bots and channels, creates webhook channels and returns the NLP result for any expression. Calls authenticate with bearer tokens generated under Settings.

A separate Tables API, with basic authentication and regional base URLs, loads and queries the built in tables that bots read and write, and whole bots export and import as files.

Sourceapi.chatlayer.ai/v1/docsread 2026-10-11

Each webhook attempt can be inspected with the agent platform's response status and body or the reason a delivery failed, which traces why an agent did not run. Template versions stay away from recipients until they are submitted, so an agent's draft can go out as a test message and be checked on a real phone first. Testing what the agent itself says or does belongs to the agent platform the customer uses.

Sourcebird.com/docs/ai/route-messages-to-an-agentread 2026-10-11

An NLP dashboard rates each language's intent model from 0 to 100 percent, shows how that rating moved since the last training, and flags intents with too few expressions, misclassified expressions and overlaps between intents.

After launch, the Train tab collects real customer messages with the suggested intent and a confidence score, and the team corrects them and adds them to the model ahead of the next release. An emulator with a debugger shows session data turn by turn, voicebots and a WhatsApp sandbox can be tried before going live, and a bad release rolls back to an earlier live version. Saved test cases that replay flows, actions or knowledge base answers are not part of the product.

Sourcedocs.chatlayer.ai/navigation/natural-language-processing-nlp/dashboardread 2026-10-11

Agents use Bird through API calls, CLI commands and MCP tools for messaging, numbers and domains. No hosted browser, remote desktop or screen control is part of the platform. Even the outbound call tool stops short of acting on its own, preparing a call that a person reviews and places from the browser.

Sourcebird.com/docs/ai/mcp-serverread 2026-10-11

Bots work through API calls, JavaScript code actions and native app integrations, and voicebots handle phone calls with speech recognition and synthesis. Knowledge base AI crawls a page or a whole domain when content is added, which builds the answer store before any conversation takes place. Chatlayer has no hosted browser, remote desktop or screen control, and Sinch's tools for agents built elsewhere work the same way, since the MCP server, Functions and Voice Relay all act through Sinch's own APIs.

Sourcedocs.chatlayer.ai/buildabot/flow-logic/dialog-state/action-bot-dialogread 2026-10-11