Bolna AI

Quick Verdict

Bolna AI combines an open-source orchestration framework with a commercial hosted platform for telephone voice agents. The canonical repository is bolna-ai/bolna, whose default branch is master. As of July 20, 2026, it was active, not archived, and explicitly seeking maintainers. The framework connects telephony, speech recognition, an LLM, and speech synthesis over streaming links: caller audio becomes text, the model or tools decide what to do, and the reply is synthesized back into the call.

Its main advantage is a replaceable provider stack and a self-hosting path, not a proprietary speech model. The core repository is MIT-licensed, while the official README says the Hosted APIs and no-code dashboard are currently closed source. Bolna markets low-latency, multilingual calling, with particular attention to Indian languages. Real conversational delay still depends on the phone network, endpoint detection, ASR, LLM first-token time, tool APIs, TTS first audio, WebSockets, and geography. A sub-600 ms product claim is not an SLA for every provider, script, and country.

Bolna can be a practical base for appointments, bounded support, lead qualification, and carefully governed outbound campaigns. It does not make a campaign lawful or safe automatically. AI identity disclosure, recording consent, do-not-call controls, opt-out, credential protection, abuse response, and human handoff must be designed as first-class product behavior.

Best For

• Teams building appointment, support, recruiting-screen, or sales-calling experiences for India and other multilingual markets.
• Engineering organizations that want a choice among Twilio, Plivo, Exotel, Vobiz, ASR, LLM, and TTS vendors rather than one fixed stack.
• Enterprises with DevOps, security, telephony operations, compliance, and human-agent capacity to operate a low-latency service.
• Teams that want to validate conversation logic with open source before choosing Hosted APIs, the dashboard, or enterprise support.

For general visual agent and knowledge-base work, compare Dify. For durable business-system automation behind a call, n8n can complement Bolna. A team without a calling-compliance owner or staffed escalation path should not run unattended outbound agents.

Key Features

• Replaceable voice stack: the open framework can orchestrate transcription, LLM, and synthesis providers. Official materials mention services such as Deepgram, OpenAI, DeepSeek, Llama-family deployments, Cohere, Mistral, ElevenLabs, AWS Polly, Cartesia, and Smallest; verify current code for the selected route.
• Telephony integration: the local repository provides Twilio and Plivo examples. Current hosted documentation also lists Vobiz and Exotel. Hosted availability must not be represented as identical open-source local support.
• Agent definition in JSON or code: configure the transcriber, LLM, synthesizer, tasks, and tool calls in a form that can be versioned and reviewed.
• Streaming and interruption handling: streaming ASR, LLM, and TTS reduce perceived delay and enable barge-in. Endpointing and recovery still require tuning by language, accent, noise, and carrier.
• Inbound, outbound, and batches: the hosted surface documents phone numbers, inbound agents, individual calls, batches, and sub-accounts for campaigns and multi-team operations.
• Tools and business systems: agents can invoke custom APIs and knowledge sources for scheduling, orders, and CRM updates. Every write must have server-side authorization, validation, and idempotency.
• Local setup: the repository’s Docker Compose example combines Bolna, a telephony service, Redis, and tunneling. External provider credentials and publicly reachable callbacks are still required.
• Human escalation: a production design should implement explicit transfer conditions, destination availability, context handoff, and a safe fallback when no human is reachable.

Use Cases

Good initial use cases are low-risk and narrowly bounded: remind an already-consenting customer about an appointment, collect a few defined fields, or answer approved FAQs. The agent should state at the beginning that it is an AI calling for a named organization, explain the purpose, disclose recording where applicable, and provide a direct way to opt out or reach a person. A realistic synthetic voice must not conceal machine identity or impersonate a particular employee. A synthetic or cloned voice may use only the speaker’s own material or recordings and media covered by explicit permission, never for unauthorized identity use, impersonation, fraud, or harassment.

Before outbound calling, validate the source of every number, marketing permission, suppression and do-not-call lists, time zones, permitted calling windows, and retry policy. Recording law varies by jurisdiction and can require one-party or all-party consent. Cross-border campaigns should identify the rules applicable to all participants and seek qualified local advice rather than applying one global banner.

Healthcare, legal, financial, collections, and employment calls create higher risk. An LLM should not make eligibility, diagnosis, credit, hiring, or binding legal decisions. Payment, account recovery, identity checks, and contractual consent need an independent secure channel. User requests, low confidence, repeated misunderstanding, complaints, tool failures, threats, self-harm signals, and regulated questions should immediately transfer to a trained human or end safely.

Security controls belong outside the prompt. Tool services must verify the caller or account, allowed fields, limits, and current state. Protect against prompt injection from speech, knowledge documents, and API responses. Do not let the model reveal system prompts, credentials, internal notes, or unrelated customer records, and never grant the LLM unrestricted database or payment privileges.

Pricing

The open-source orchestration repository is free under MIT, but self-hosting requires servers, Redis, a telephony provider, ASR, an LLM, TTS, numbers, monitoring, and operations. On July 20, 2026, the hosted pricing page presented pay-as-you-go credits, Pilot, and Enterprise options. Its fixed-volume illustration used 6 cents per minute with 30-second billing pulses and volume discounts, and a small signup credit was advertised. Prices can change, so use the live checkout and contract instead of preserving an older 2-cent platform-fee claim.

There is no single voice-agent per-minute cost. Total cost combines Bolna hosted usage or self-host infrastructure, call termination and phone-number charges, ASR audio minutes, LLM tokens and tool calls, TTS characters or audio, recording storage, observability, retries, and human-transfer time. BYOK may remove a platform markup for connected providers, but the provider still bills the account. A pilot should measure connected minutes, voicemail and silence, failed retries, tokens, TTS usage, transfer duration, and completed outcomes. Cost per successful task is more meaningful than a headline call-minute rate.

MIT covers the Bolna repository code, not every connected ASR, LLM, TTS checkpoint, cloud API, voice, phone number, or dataset. Review each provider and self-hosted model under its own model card, service terms, and data-processing agreement. Permission to call an API does not imply permission to retain recordings for training or to present a synthetic voice as a real person’s identity.

Pros

• An inspectable, extensible open core provides a self-hosting option and reduces dependence on one all-in-one voice platform.
• Telephony, ASR, LLM, and TTS choices can be routed by language, geography, quality, latency, and cost.
• Streaming, batch calls, tools, and hosted APIs cover most of the end-to-end voice-agent lifecycle.
• The product has visible investment in Indian language use cases and provider choices relevant to that market.
• A team can begin with an open local proof of concept and later evaluate Hosted or Enterprise support.
• JSON and code configuration can support change review, testing, and controlled deployment better than an opaque prompt-only agent.

Cons

• Hosted APIs and the no-code UI are closed source, and their providers and features should not be conflated with the repository.
• End-to-end latency crosses several external layers; marketing figures do not replace peak-load and target-region testing.
• Several bills compose the stack, and failed calls, retries, recordings, numbers, and human transfers are easy to omit from forecasts.
• Self-hosting still requires public telephony callbacks, Redis, secrets, resilience, monitoring, patching, and incident response.
• Telemarketing permission, do-not-call, recording, AI disclosure, cross-border data, and industry rules are organizational duties, not framework features.
• Tool calling and spoken prompt injection can create unauthorized business actions unless permissions are enforced in backend services.

Alternatives

FAQ

Is all of Bolna open source?

No. The core orchestration repository uses MIT. The official README says the Hosted APIs and no-code dashboard are currently closed source. Evaluate open code, hosted product, and enterprise services separately.

Which telephony providers are officially supported?

The local repository clearly includes Twilio and Plivo setup examples. Current hosted provider documentation also lists Vobiz and Exotel. Verify inbound, outbound, country, number, and bidirectional-streaming support for the exact product surface.

Does sub-600 ms mean every reply arrives within 600 ms?

No. It is a low-latency positioning claim, not a universal guarantee. Measure P50, P95, and P99 across endpointing, ASR, model, tools, TTS, network jitter, carrier playback, and interruption recovery in the target region.

What is the actual cost per minute?

Add platform or infrastructure, telephony, numbers, ASR, LLM, TTS, recordings, monitoring, retries, and human transfer. Hosted pricing covers only part of the stack, and BYOK usage still generates vendor charges.

Must an AI call disclose itself and recording?

The agent should clearly disclose its AI identity and organization at the start. Recording consent depends on the jurisdictions of the participants and can require one-party or all-party agreement. Outbound campaigns also need permission, opt-out, suppression, and calling-window controls.

When should a call transfer to a human?

Transfer on user request, identity or payment issues, regulated or high-risk topics, low confidence, repeated failure, complaints, threats, self-harm signals, tool errors, or any question not grounded in approved information.

How should prompt injection and tool abuse be controlled?

Do not give the LLM direct high privilege. Backend tools must validate identity, fields, limits, and state and apply least privilege, idempotency, approvals, rate limits, audit logs, and anomaly alerts. Treat spoken and retrieved content as untrusted input.

Bottom Line

Bolna is a flexible way to assemble telephony, ASR, an LLM, TTS, and business tools into a streaming voice agent while retaining an open self-hosting option. Selection should depend on target-region P95 latency, cost per completed outcome, carrier and provider failure behavior, and the quality of human takeover, not only a natural demo. Validate number and recording rules first, then pilot a low-risk workflow at small volume. Scale only when AI identity disclosure, consent and opt-out, secret isolation, tool authorization, recording protection, abuse response, and human escalation are implemented and auditable.