mcp-telegram

MCP server for Telegram — let AI assistants interact with your Telegram account

Features

• Chat Management: List, search, mute/unmute chats, organise into folders

• Messages: Read, search, inspect context, send, draft, schedule, link-resolve, and backup messages

• AI Summarisation: Summarise chat conversations using multiple LLM providers

• Secure storage: separate versioned Keychain items on macOS; atomic 0600 session/config files in a 0700 state directory on Linux and Windows

• Two transports: local stdio (single account) or remote streamable HTTP with an embedded OAuth 2.1 server and per-user Telegram QR login — multi-user and deployable to Cloud Run (see Remote (HTTP) Mode)

Related MCP server: Better Telegram MCP

Installation

go install github.com/tolmachov/mcp-telegram@latest

Or build from source:

git clone https://github.com/tolmachov/mcp-telegram.git
cd mcp-telegram
make

Setup

1. Get Telegram API Credentials

1. Go to my.telegram.org/apps

2. Create an application

3. Copy api_id and api_hash

2. Configure Environment

Store credentials (one versioned Keychain item per value on macOS; one atomic 0600 file per value in a 0700 state directory on Linux/Windows):

mcp-telegram config set api-id 123456789
mcp-telegram config set api-hash abcd1234efgh5678

Or export the current environment names explicitly:

export MCP_TELEGRAM_API_ID=123456789
export MCP_TELEGRAM_API_HASH=abcd1234efgh5678

3. Login to Telegram

mcp-telegram login --phone +1234567890

You'll be prompted for the verification code sent by Telegram and, when the account has two-step verification enabled, its 2FA password.

4. Configure MCP Client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "telegram": {
      "command": "mcp-telegram",
      "args": ["run"],
      "env": {
        "MCP_TELEGRAM_API_ID": "your_api_id",
        "MCP_TELEGRAM_API_HASH": "your_api_hash"
      }
    }
  }
}

Claude Code

claude mcp add telegram -- /path/to/mcp-telegram run

Set process environment variables or pass them through your MCP host. Local .env files are not loaded automatically.

Available Tools

The full stdio surface exposes 29 tools (full over HTTP exposes 28 because filesystem backup is local-only). Messages are identified by opaque string handles ("42" for regular, "s:42" for scheduled) — copy them back verbatim from tool outputs to follow-up calls, never parse or construct them manually.

If your client shows a single tool called TelegramLoginRequired and none of the tools below, the server is up but Telegram is not authorized — see When Telegram Isn't Authorized.

The folder edits (CreateFolder, AddChatsToFolder, RemoveChatsFromFolder) report chats that cannot be resolved in skipped and apply the rest; any other resolve failure, such as a flood wait or a network error, fails the whole call so it can be retried.

A call that gets only part of its work done — a batch stopped early, a listing or summary that is incomplete, a backup that saved what it fetched — is not reported as a failure: its result carries a warning saying what is missing and why.

Pagination, dates, and identifiers

• GetMessages, SearchMessages, and GetReplies accept filters only on the first call. For the next page, copy next_cursor into cursor and omit every other field. GetForumTopics follows the same cursor-only continuation rule.

• before_message_id is an optional first-page anchor, distinct from the continuation cursor. Scheduled handles cannot be pagination anchors.

• Cursors are opaque and versioned. Do not decode, modify, or construct them; cursors from older or newer incompatible schemas are rejected.

• Message date windows are consistently [from_date, to_date): from_date is inclusive and to_date is exclusive. Both are RFC3339; an empty or reversed window is a validation error. To include a calendar day, use midnight of the following day as to_date.

• IDs returned in message DTOs and pinned resources use the same opaque string handle format. Pass them back exactly as returned.

Server Variants

The server implements experimental server variants (SEP-2053): one server that offers several selectable capability sets. A client sends hints during initialize, the server ranks the variants, and the client picks one. Clients that don't understand the extension transparently get the full variant, so nothing changes for them.

Pin a single variant with --variant (or MCP_VARIANT) for clients that can't negotiate — e.g. --variant research exposes only the read-only subset:

mcp-telegram run --variant research

Leave it unset to expose all three and let the client choose.

Note on pinned-chat resources. In multi-variant mode (no --variant), pinned-chat resources are exposed on every variant and kept fresh by a single poller, but proactive resources/list_changed notifications are not delivered — the variants proxy can't forward the background watcher's notifications (an upstream library limitation). Clients pick up pin changes on their next resources/list. Pin a single --variant to restore live notifications.

Available Resources

Pinned chat resources are created dynamically for each pinned chat and refreshed in the background; clients receive resources/list_changed when the set changes (except in multi-variant mode — see the note under Server Variants).

Available Prompts

3 parameterized prompts that MCP clients expose as slash-commands or quick actions.

Prompt Examples

Here are some example prompts you can use with AI assistants:

Message Management

• "Check for any unread important messages in my Telegram"

• "Summarise all my unread Telegram messages"

• "Read and analyse my unread messages, prepare draft responses where needed"

• "Check non-critical unread messages and give me a brief overview"

• "Find messages mentioning 'invoice' in my work chat from last week"

• "Open the context around this Telegram link: https://t.me/example/123"

Organisation

• "Analyse my Telegram dialogs and suggest a folder structure"

• "Help me categorise my Telegram chats by importance"

• "Find all work-related conversations and suggest how to organise them"

Communication

• "Monitor specific chat for updates about [topic]"

• "Draft a polite response to the last message in [chat]"

• "Check if there are any unanswered questions in my chats"

• "Resolve this Telegram message link and show me the thread context"

Backup & Export

These examples apply only to local stdio full/compact mode; HTTP and the research variant never expose filesystem backup:

• "Backup my conversation with [contact] to a file"

• "Export the last week of messages from [group]"

• "Backup media-only updates too so nothing is silently skipped"

Chat Summarisation

The SummarizeChat tool supports multiple LLM providers:

• sampling (experimental): Uses the MCP client's LLM via MCP Sampling. Only works with clients that support sampling: VS Code, fast-agent, Continue. Does NOT work with Claude Desktop or Claude Code.

• ollama: Local LLM via Ollama - no API key required

• gemini: Google Gemini API

• anthropic: Anthropic Claude API

max_messages defaults to 2000 and cannot exceed 10000. Fetching is bounded and paged in batches of 100; if a later Telegram page fails, the tool summarises the messages already fetched and marks the result as degraded. Every response includes messages_processed and truncated, with partial and warning when applicable.

Chat messages are serialised as untrusted JSON data, separate from the system instructions, goal, and previous rolling summary. The server explicitly marks Telegram content as untrusted and tells the provider not to execute instructions found inside it. Sampling sends selected text to the MCP client's LLM; Gemini and Anthropic send it to their APIs; Ollama sends it to the configured URL, which may itself be remote. Providers may log, retain, or bill for content under their own policies. Retryable 429/5xx responses are attempted at most three times with backoff and Retry-After, without exceeding the MCP request deadline.

Configure via environment variables:

MCP_SUMMARIZE_PROVIDER=ollama  # or: sampling, gemini, anthropic
MCP_SUMMARIZE_MODEL=           # provider-specific model name

Summarisation is optional, so invalid settings — an unknown provider, a missing Ollama URL, a non-positive batch size, an API key that is not set or cannot be read from the Keychain / config store — never stop the server: they are logged as a warning at startup, every other tool works, and SummarizeChat answers each call with the exact startup error and the fix. The settings are read once, so after fixing them reconnect the server (stdio) or restart it (HTTP).

Commands

# Run MCP server (used by MCP clients)
mcp-telegram run

# Login to Telegram
mcp-telegram login --phone +1234567890

# Logout and delete session
mcp-telegram logout

# Securely store config values (macOS Keychain / file on Linux)
mcp-telegram config set api-id 123456789
mcp-telegram config set api-hash abcd1234

# List stored keys
mcp-telegram config list

# Delete a stored value
mcp-telegram config delete api-id

Allowed keys: api-id, api-hash, anthropic, gemini.

logout always closes the local client and removes the local session. If Telegram is offline, the command reports the remote logout failure together with any local cleanup failure instead of leaving the local credential behind.

For storable credentials, configuration resolves in this priority order: CLI flags → process environment → secure store → defaults. Other runtime settings resolve as CLI flags → process environment → defaults. The binary never reads .env files and never mutates the process environment.

Configuration Options

.env.example is documentation only. Export values in the actual server process, pass flags, or use config set for supported credentials. On Cloud Run, K_SERVICE causes the application to validate and consume the platform's injected PORT; do not put $PORT inside MCP_HTTP_ADDR because Cloud Run does not expand environment-variable references there.

Remote (HTTP) Mode

By default the server speaks the stdio transport: one local process, one Telegram account, driven by a desktop MCP client. It can instead run as a long-lived remote server over the MCP streamable HTTP transport, so a single hosted deployment serves many users — each authenticating with their own Telegram account — over the network.

How authorization works

Authorization is pure Telegram — no bots, no external identity provider. The server embeds a small OAuth 2.1 authorization server (Dynamic Client Registration + PKCE); when a client connects it opens the authorization page, which shows a QR code. You scan it with the Telegram app (Settings → Devices → Link Desktop Device), and the resulting MTProto session both proves who you are and becomes your working session (a 2FA-password prompt appears if your account has one). Only Telegram user ids listed in MCP_AUTH_ALLOWED_USERS may complete the login — everyone else is rejected after the scan and their session is discarded.

Each authorization is an independent session: it gets its own encrypted record in the selected backend and its own MCP server assembly, so one account can be logged in from several clients at once without them contending. The server keeps up to four assemblies per account and 25 globally. An account actively switching among more than four authorizations is not rejected, but its least-recently-used assembly is evicted and rebuilt on demand. Sessions are encrypted at rest with AES-256-GCM under a split key: the key is derived from both MCP_AUTH_TOKEN_KEYS and a random per-session key that lives only inside the client's OAuth access/refresh token (never stored server-side). As a result, an at-rest dump of the bucket plus the secret manager cannot, on its own, decrypt a session — a live token is also required. (Trade-offs, stated plainly: this does not protect against a compromise of the running server's memory. And because the per-session key share travels inside the token, an attacker who holds the bucket and the master key and captures one access token can decrypt that single session — and thereby extract its persistent MTProto auth key, which outlives the ≤55-minute token and keeps working until the session is revoked or logged out. That exposure is confined to the one captured session; other sessions stay protected. TLS, Cache-Control: no-store on token/revoke responses, and never logging tokens narrow the capture window.) Old session formats are intentionally unreadable.

Session lifecycle is managed automatically: POST /revoke (RFC 7009) with an access or refresh token durably marks that authorization revoked (a tombstone stored where the Telegram client never writes) and deletes its session. The refresh grant dies immediately and stays dead — even if a still-live client re-stores the session object, the refresh check consults the tombstone, so revocation cannot be undone by a resurrected blob. An already-issued access token remains valid until it expires (≤55 minutes): revocation reliably stops renewal, matching the standard short-lived-access / revocable-refresh model. Revocation does not terminate the Telegram-side device authorization (it stays in Settings → Devices until Telegram expires it). Sessions abandoned without revocation — and old tombstones — are reclaimed by a background sweep once older than the refresh-token TTL plus a day, past which no refresh token for the session can still be valid.

Authorization codes are durable single-use grants. Refresh tokens rotate by atomic family generation; reuse of an older generation revokes the entire family, Telegram session, and live pooled client. Backend failures return 503 without consuming or rotating a grant.

The HTTP boundary limits MCP bodies to 1 MiB, headers to 32 KiB, and concurrent requests (including SSE streams) to 128. Header/read/idle timeouts are 10s/30s/60s; MCP sessions expire after 10 minutes. Creating sessionless MCP sessions is limited per Telegram user to 10 per minute with burst 3. By default MCP_AUTH_TRUSTED_PROXY_HOPS=0, so forwarded-address headers are ignored; configure a positive value only when the complete trusted proxy chain is known. Origin protection remains enabled around the authenticated MCP endpoint to block cross-origin and DNS-rebinding requests.

Try it locally

# Issuer on loopback may use plain http:
mcp-telegram run --transport http --http-addr 127.0.0.1:8080 \
  --auth-issuer-url http://localhost:8080 \
  --auth-allowed-users 123456789 \
  --auth-token-key "$(head -c 32 /dev/urandom | base64)" \
  --auth-session-dir ~/.local/state/mcp-telegram/http-sessions

Then point an MCP client at http://localhost:8080/ and complete the browser flow. HTTP cannot start without OAuth configuration and exactly one durable session backend. Use * as the sole entry in --auth-allowed-users to allow any Telegram account (the deployment is then only as private as its URL; * cannot be combined with specific ids).

Deploy to Google Cloud Run

The production target is Cloud Run: one container, scale-to-zero, secrets from Secret Manager, sessions in a GCS bucket. In outline:

1. Create a bucket for the encrypted per-user sessions and a dedicated service account with roles/storage.objectAdmin on just that bucket.

2. Store two secrets in Secret Manager — the Telegram api_hash and a freshly generated 32-byte MCP_AUTH_TOKEN_KEYS (head -c 32 /dev/urandom | base64) — and grant the service account roles/secretmanager.secretAccessor on them. The token key never leaves Secret Manager in plaintext. Losing it logs everyone out (sessions become undecryptable). Leaking it alone no longer exposes split-key sessions — those also need a live client token — but treat it as highly sensitive regardless: it can mint tokens and decrypt current sessions when combined with their per-authorization token key.

3. Deploy with gcloud run deploy --source ., wiring the non-secret env from deploy/cloudrun.yaml.example, the two secrets via --set-secrets, and --max-instances=1 (mandatory: two instances loading the same MTProto session trip Telegram's AUTH_KEY_DUPLICATED and forcibly log every user out).

4. Set MCP_AUTH_ISSUER_URL to the service URL and redeploy — tokens and session encryption are bound to the issuer value.

Do not set MCP_HTTP_ADDR=:8080 for Cloud Run. The platform injects PORT, and the application binds to :$PORT when K_SERVICE is present. An explicitly set MCP_HTTP_ADDR remains an override for non-Cloud-Run environments.

Copy-paste commands, the exact IAM bindings, the token/revocation model, and how to connect claude.ai / Claude Desktop / Claude Code are in deploy/README.md.

Destructive Actions

DeleteMessages, DeleteFolder, LeaveChat, and ForwardMessage require an explicit confirm: true. Missing confirmation always fails closed, regardless of the client's elicitation capabilities.

When Telegram Isn't Authorized

Telegram sessions expire and can be revoked from Settings → Devices → Active sessions on any of your devices. When that happens at startup — or when api-id/api-hash are missing — the server does not fail its MCP connection. Over stdio it comes up in login-required mode:

• the host shows the server as connected, exposing a single tool named TelegramLoginRequired and no Telegram tools at all;

• the server instructions tell the model Telegram is unavailable and what the fix is, so the first Telegram request you make gets answered with the real reason instead of a generic failure;

• calling TelegramLoginRequired re-checks the live state and reports not_configured, login_required, check_failed, or authorized_pending_reconnect — the last one meaning you logged in elsewhere and only need to reconnect the server.

The fix is the usual one. Use the path your host actually launches — the binary is typically wired in by absolute path and not on $PATH, which is why the tool's fix_command field hands back a paste-ready command with the real path already filled in:

/path/to/mcp-telegram login --phone +1234567890
# then reconnect the MCP server (in Claude Code: /mcp → select this server → Reconnect)

This is deliberate. A stdio server that refuses its MCP connection is rendered by hosts as a bare "failed" entry with the reason buried in a log file, indistinguishable from a bad binary path — so the diagnosis is delivered as server content instead, which is the only channel a stdio server has. Two paths keep fail-fast behaviour, because they have no MCP peer to tell:

• the HTTP transport — the process exits non-zero so Cloud Run sees an unhealthy start rather than a listener it can never serve;

• stdin attached to a TTY, i.e. you ran mcp-telegram run yourself — the message goes to stderr and the process exits, which is what makes that command usable as a manual smoke test.

Note the boundary is the TTY, not "a human started it": with stdin piped or redirected from a file, mcp-telegram run serves login-required mode and exits 0 on EOF. Scripted health checks should call mcp-telegram login/config list, or assert on the tool list, rather than on run's exit status.

If Telegram revokes the session while the server is running, the first call the account's home data centre refuses stops the client for good: every tool call then answers with the reason and fix instead of reaching Telegram, and a call the stop cut short gets them appended to its own result. The host stays connected, and after mcp-telegram login a reconnect of the server loads the tools again. A client that stops for any other reason is reconnected in-process on the same session, and calls meanwhile answer that it is and to retry what they did not complete. That includes file downloads and statistics calls refused by another data centre, whose refusal says nothing about the session: the client reconnects, and the reconnect's own session check has the home data centre judge it.

In remote (HTTP) mode a dead per-user session — whether Telegram refuses it at connect time or the home data centre refuses a call — is deleted by the request that finds it dead. That request is answered with 401 plus a WWW-Authenticate challenge, which sends the MCP client back through OAuth and its QR login to mint a fresh session — no restart, no CLI. A client that stops for any other reason, such as a dropped connection or another data centre refusing the session, is reconnected on the same session at the next request, whose session check has the home data centre judge it.

Session, Config, and Backup Storage

Local stdio storage is intentionally versioned and does not read the previous formats:

• macOS: Telegram session data and every stored config key are independent generic-password items under the mcp-telegram.v2 Keychain service. Reads go directly to Keychain; there is no process cache or aggregate JSON blob.

• Linux/Windows: the state root is $XDG_STATE_HOME/mcp-telegram, or $HOME/.local/state/mcp-telegram when XDG_STATE_HOME is unset. The Telegram session is session-v2.bin; config values are separate files under config-v2/. The directory is forced to 0700, files are 0600, and writes use fsync plus atomic replacement so concurrent updates to different keys do not overwrite one another.

• Backups: the default stdio backup root is <state-root>/backups. An operator may replace it with --allowed-paths or MCP_TELEGRAM_ALLOWED_PATHS. Client-provided MCP roots never expand the filesystem allowlist. Targets are checked after symlink resolution and files are atomically replaced. HTTP and research do not initialise or expose this facility.

The Linux/Windows local session and config files are plaintext despite their restrictive permissions. A copied session-v2.bin can grant access to the Telegram account: keep the OS account trusted, use an encrypted filesystem (LUKS/BitLocker), and do not sync the state directory to an unencrypted cloud backup.

HTTP mode does not use the local stdio session. It requires exactly one backend (MCP_AUTH_SESSION_BUCKET or MCP_AUTH_SESSION_DIR) and stores encrypted, per-authorization objects under sessions-v3/, durable revocation tombstones under revoked-v3/, and OAuth state under oauth-v3/. Session decryption requires both the server token key and the per-authorization key carried inside the sealed client token.

Breaking Upgrade

This release deliberately provides no migration or compatibility layer:

• previous env names and the removed --auth/unauthenticated HTTP modes are not accepted;

• the old aggregate Keychain blob, config JSON, local session path, HTTP tokens, HTTP storage prefixes, and old opaque cursors are not read;

• every HTTP authorization and affected local installation must be configured with the current MCP_* names and complete a new Telegram login/QR login.

Before deployment, record the previous image digest and retain the previous storage long enough to roll back. Delete old HTTP objects only after the new release passes the real-Telegram and Cloud Run smoke described in deploy/README.md. A rollback must restore the previous binary, environment, and storage prefixes together.

Development

make build            # build the binary with version metadata
make lint             # golangci-lint
make test             # unit tests
make test-integration # end-to-end tests (need a real account + TEST_* vars)

# Production acceptance checks used by CI
go test -race ./...
./scripts/check-coverage.sh
go vet ./...
govulncheck ./...
go mod verify

Unit tests run without credentials. The integration suite in test/ is behind the integration build tag and drives the server over a real stdio pipe using a second MCP implementation (mark3labs/mcp-go) as the client, to catch wire-level interop issues. It reads the TEST_* variables documented in .env.example (e.g. TEST_CHAT_ID, TEST_GROUP_ID) and skips any test whose variable or Telegram credentials are unset.

CI runs ordinary tests on Linux and macOS, the race detector, package coverage gates, vet, golangci-lint, govulncheck, a Windows cross-build, and a pinned Docker build followed by a HIGH/CRITICAL vulnerability scan. The release image uses a checksummed Go 1.26.8 toolchain and Alpine 3.22 with security updates.

License

MIT