call4me

https://call4.me

Your coding agent (Claude Code, Codex, Claude Desktop, ChatGPT) gets one new tool: make a phone call. Prepaid credits from $10 via Stripe Checkout, reloading monthly by default; one MCP URL with the key in it.

How a call works

agent ──MCP──▶ worker ──POST /v2/calls──▶ Telnyx ──PSTN──▶ business
                  ▲                          │ media stream (PCMU, bidirectional RTP)
                  │ webhooks                 ▼
                  └──────────────── VoiceSession (Durable Object) ◀──WS──▶ GPT-Live (gpt-live-1, audio/pcmu 8k)
                                                                           └─ back office (Responses model):
                                                                              end_call · ask_user · press_digits

• src/server/voice/prompt.ts: what the caller is told. The product lives here: no opening disclosure, no recording notice, no end-of-call read-back; honest if sincerely asked whether it's an AI.
• src/server/voice/session.ts: the audio relay and back-office functions.
• src/server/services/dialer.ts: placing calls, per-account numbers, answering callbacks.
• src/server/services/numbers.ts: the account's call4me numbers, and the user's own numbers verified to call from. Telnyx keeps one verified-number list for our whole carrier account, so verified_numbers ties each to the one call4me account that submitted a fresh code for it; a number Telnyx already lists as verified is never handed to another account. A call from an own number (from) only reaches businesses in its country, and still rings the user from a call4me number.
• src/server/mcp/server.ts: the MCP tools.

The app does not request recording when dialing or answering. Existing recordings stored by Telnyx can be played or opened from the signed in call page after a call ends, or retrieved using call4me_get_recordings({ call_id }) over MCP, or GET /api/calls/{call_id}/recordings with a call4me API key or OAuth token in the Bearer header. Both check account ownership before contacting Telnyx and return fresh download links without storing them. Empty results can mean processing is pending or no recording was saved. Links may expire and should only be shared with the account owner.

Trunk recordings use call leg IDs. A separate provider ID table receives these from verified webhooks; existing calls resolve them from an exact control ID match in Telnyx webhook history. The app never guesses a match from phone numbers or timestamps. If provider history has expired and no mapping was saved, only recordings directly associated with the control ID can be found.

The recording lookup never contacts the voice session or changes prompts, live tools, audio, or Telnyx recording settings. Existing call status and transcript responses are unchanged.

The recording API's OpenAPI document is served at /api/openapi.json. Run npm run openapi:generate to regenerate openapi.json from the shared MCP and HTTP schemas. CI checks the generated artifact for drift. Run npm test, npm run check, npm run lint, and npx wrangler deploy --dry-run before deploying through git.

Menu recovery tracks fresh phone prompts separately from transcript timestamps and remembers accepted keypad submissions. Repeated menus, invalid input, and recorded instruction dead ends can restart the existing back office after the recording goes quiet. It waits for pending tools and their continuations, avoids interrupting holds or a person answering, and asks the model to choose a supported route from the heard options rather than blindly replaying digits.

npm test includes provider message replay tests for recovery and its timing races. To evaluate the configured back office model against synthetic menu scenarios, run npm run test:menu:live with OPENAI_API_KEY (or a local .dev.vars). This opt-in evaluation calls OpenAI, inspects the proposed tool calls, and never executes a tool or places a phone call.

Setup

1. Telnyx: a Call Control application (its id is TELNYX_CONNECTION_ID) with webhook URL https://<host>/webhooks/telnyx, an API key, and the account's webhook public key. Each account's number is bought on its first call and attached to that application.
2. OpenAI: an API key with GPT-Live access.
3. Stripe: webhook endpoint https://<host>/webhooks/stripe for checkout.session.completed, checkout.session.async_payment_succeeded, invoice.paid, customer.subscription.updated, customer.subscription.deleted, charge.refunded. The blog's supporter subscriptions arrive on the same endpoint, tagged app=callbay, kind=supporter (see src/server/services/supporters.ts).

Credits

Everything is prepaid. Loads start at $10 and, by default, the same amount reloads monthly (a Stripe subscription; invoice.paid adds the credits). A call holds its maximum cost before it dials, in one conditional insert so parallel calls can't overspend, and settles to the real talk time when it ends. 4. Secrets (wrangler secret put): see src/server/env.d.ts.

Two Workers

Call sessions run in their own Worker, call4me-voice (wrangler.voice.jsonc, src/server/voice/worker.ts), which also takes Telnyx's media streams on voice.call4.me. Deploying a Worker resets the durable objects it defines, so with the sessions split out the site deploys freely without dropping a call. CI deploys the voice Worker with scripts/deploy-voice.sh, which skips an unchanged bundle and otherwise waits until no call is up, deploys, and holds new calls until the new version is serving plus a minute for Cloudflare to retire the old one. Never deploy it or change its secrets any other way: either restarts every live call. The voice Worker's secrets are OPENAI_API_KEY, TELNYX_API_KEY, TELNYX_CONNECTION_ID, RAINDROP_WRITE_KEY and STREAM_SECRET (the same value as the site's): npx wrangler secret put <NAME> -c wrangler.voice.jsonc.

Crawlers

/robots.txt allows every crawler on every path, including AI search, answers, and training. The wildcard covers current and future crawlers without maintaining a named allowlist. Cloudflare's desired crawler settings live in cloudflare.crawlers.json. Run npm run crawlers to audit them, or npm run crawlers -- --apply to apply and verify them. The command requires CLOUDFLARE_API_TOKEN with Zone Read, Zone Settings Read and Write, and Bot Management Read and Write for call4.me. Any unavailable or mismatched setting fails the command. Separately audit AI Crawl Control and firewall rules for crawler blocks. Cloudflare's automatic DDoS protection remains active.

If an account token can manage tokens but lacks the permissions on this zone, use npm run crawlers -- --apply --authorize to add the necessary permissions for call4.me. It discovers permission IDs from Cloudflare and preserves existing permissions and restrictions. The token editor calls Write permissions Edit. Account tokens may be unsupported by some bot products; use a user token from My Profile if Cloudflare rejects a correctly scoped account token. To manage another site's committed policy, pass --policy /path/to/cloudflare.crawlers.json.

Public page delivery

npm run build:site renders public HTML and Markdown from the existing Hono views into dist/site, copies the public assets, and fingerprints CSS and JavaScript. Wrangler runs this build automatically before local development, dry runs, and deployments. CI also builds before checking the generated route manifest and tests. No database or credentials are used to generate public pages. Paid articles are excluded from the static manifest.

src/server/static-entry.ts serves these documents through Cloudflare's cached asset binding without initializing the application or reading D1. CSS, images, and scripts go straight to static asset delivery. Native page transitions progressively enhance ordinary navigation and honor reduced motion. Audio keeps the byte range handler needed for seeking.

Public content is identical for every viewer. Account navigation, counts, and escaped blog comments load after the page paints from private, uncached live endpoints. Search, top sorting, subscription feedback, authenticated MCP setup, accounts, payments, OAuth, and all writes still use the existing server. ?live=1 opens the dynamic page when JavaScript is disabled or live comments cannot load. First touch cookies remain per visitor; a paid Reddit landing records its attribution before issuing the visit cookie.

Blog

/blog is markdown files in src/content/blog (one per post, listed in index.ts, headline image at public/static/blog/<slug>.svg), with an Atom feed, likes, comments, an email newsletter (sent from /admin/blog, for the ADMIN_EMAILS accounts), and paid posts for monthly supporters.

Related articles use local CPU embeddings, ranked within editorial subjects from scripts/related-topics.ts. After adding an article or changing its SEO title, tags or subject, run npm run blog:related, review the matches in src/content/blog/related.generated.ts, and commit the index. The first run downloads the pinned EmbeddingGemma model; later runs use ~/.cache/call4me/transformers. Article text stays local. The site serves the generated index without model inference. Articles can have fewer than three recommendations when matches fall below the similarity threshold. npm test checks freshness without downloading a model.

Self-hosting

Fork it, create your own D1 database (wrangler d1 create callbay) and put its id and your own hostname in wrangler.jsonc.

Local: npm run db:migrate:local && npm run dev with a .dev.vars holding the same secrets.

AI monitoring

Raindrop monitors the live conversation (callbay_voice_call), back office model runs (callbay_back_office), and the recap after hangup (callbay_call_recap). Each event uses the account ID as its user ID and the call ID as its conversation ID. Model names, timing, status, and back office tool spans are included. No phone audio or recording URLs are uploaded.

Add RAINDROP_WRITE_KEY to .dev.vars locally and configure the same key as a Cloudflare Worker secret with npx wrangler secret put RAINDROP_WRITE_KEY. Optional RAINDROP_PROJECT_ID selects a project slug; unset uses the write key's default project. Without a write key, monitoring is disabled. .dev.vars.example lists these optional settings.

AI inputs and outputs use Raindrop's PII redaction, with Callbay's existing masking for per call secrets applied first. Tool arguments and results appear in redacted back office AI output. Tool spans contain names, timing, and generic failure status, keeping sensitive payloads out of unredacted trace attributes. Account names and emails are not sent as user traits. SDK PII redaction is pattern based and does not guarantee removal of every sensitive detail.

Monitoring runs separately from live audio and tools. Terminal events and queued tool spans are flushed within the Worker or Durable Object lifetime. An unfinished back office run at hangup is marked interrupted; later tool results do not reopen the closed monitoring client.

After a call, check the three event names in Raindrop, with actual inputs and outputs, the matching account and call IDs, and tool names and durations. A call that never invokes a back office model has no back office event. Feedback signals, audio attachments, and agent self diagnostics are not instrumented because the current call flow has no corresponding feedback controls or diagnostics tools.

For investigation, connect the Raindrop MCP server and install the investigation skill with npx skills add raindrop-ai/skills --skill raindrop-investigate. Connect Slack in Raindrop for alerts. Create a Raindrop account if needed.

License

MIT