This ZeroClaw Ollama setup connects a locally running model to a Telegram bot. I tested the complete route on 13 October 2026: a Telegram message reached ZeroClaw, Ollama generated the answer, and the bot returned “2 + 2 = 4.” The successful exchange is shown in Step 5.
Tested versions: ZeroClaw 0.8.5, Ollama 0.35.0, and qwen3:0.6b, on an Apple M3 Mac. The small model download was 522 MB. This is a basic chat test, not a benchmark or a demonstration of reliable autonomous tool use. Linux installation notes are included, but the screenshots come from the macOS test.
Local model, cloud messaging: Ollama runs the model on your computer, but Telegram messages still pass through Telegram’s servers. A Telegram bot is not a Secret Chat and this setup is not fully offline or end-to-end encrypted. Use ZeroClaw’s terminal chat if you want to avoid Telegram transport. See Telegram’s privacy explanation and Ollama’s local-data explanation.
The message path is: Telegram app → Telegram cloud/Bot API → ZeroClaw on your computer → local Ollama → Telegram cloud → your chat. Your computer and both local processes must remain running for the bot to respond.
Before you start#
You need a terminal, enough memory for your chosen model, internet access for installation and Telegram, and a Telegram account. This guide uses a fresh ZeroClaw configuration. Back up an existing configuration before changing its agent or channel settings.
The commands below use the normal zeroclaw executable and its default configuration directory, ~/.zeroclaw. For the temporary test, I used a separate folder with zeroclaw --config-dir /path/to/test-config. If you do that too, include the same --config-dir on every ZeroClaw command.
Screenshots have arrows in empty space. Command screenshots are labeled views of recorded output, rather than pictures of an installation wizard. Telegram screenshots are from the real test, cropped and annotated; private identifiers and the pairing code are redacted.
Step 1: Install ZeroClaw and prepare Ollama#
Get the binary for your operating system and architecture from the official ZeroClaw v0.8.5 release. Extract the archive, compare its SHA-256 checksum with the release’s SHA256SUMS, and put the zeroclaw executable in a directory on your PATH. The prebuilt binary avoids a Rust source build.
For example, on an Apple Silicon Mac, after downloading the release archive and SHA256SUMS into the same folder:
shasum -a 256 zeroclaw-aarch64-apple-darwin.tar.gz
# Compare the result with this archive's entry in SHA256SUMS.
tar -xzf zeroclaw-aarch64-apple-darwin.tar.gz zeroclaw
mkdir -p "$HOME/.local/bin"
install -m 755 zeroclaw "$HOME/.local/bin/zeroclaw"
export PATH="$HOME/.local/bin:$PATH"
zeroclaw --version
Add that PATH setting to your shell profile if needed. Linux users should choose the matching Linux archive from the same release; the macOS binary will not run on Linux. For other installation options, use the official installation guide.
Install Ollama from the official download page. On Linux, the documented installer is:
curl -fsSL https://ollama.com/install.sh | sh
If Ollama is not already running, start it in a separate terminal and leave it open:
ollama serve
Do not start a second server if the Ollama app or system service already owns port 11434. Then download the same model used here:
ollama pull qwen3:0.6b
ollama --version
ollama list

Checkpoint: the exact model tag appears in ollama list. The screenshot records the versions and model from this test.
Ollama also supports cloud models. This guide uses the downloaded local tag and a loopback endpoint. If you want to disable Ollama’s cloud features entirely, set OLLAMA_NO_CLOUD=1 for the Ollama server and restart it, as described in the Ollama FAQ. That does not disable Telegram’s cloud.
Step 2: Create an agent and test the local model#
For ZeroClaw 0.8.5, use Quickstart. The older onboard flow shown in the previous version of this article is deprecated.
zeroclaw quickstart --model-provider ollama --model qwen3:0.6b --agent guide
In the selector, choose Locked Down for the risk profile, none for memory, and open Channels, then choose Done without adding a channel yet. Leave peer groups empty. Select Create agent. In this fresh test, Quickstart created the provider alias ollama.default and runtime profile unbounded.
Apply the settings used for the successful test:
zeroclaw config set providers.models.ollama.default.uri http://127.0.0.1:11434
zeroclaw config set providers.models.ollama.default.think false
zeroclaw config set providers.models.ollama.default.num_ctx 4096
zeroclaw config set runtime_profiles.unbounded.agentic false
The last setting makes this a chat-only walkthrough. A tiny model is enough to check the connection; test a more capable model separately before enabling agent tools. If you use an existing setup, inspect its aliases first and avoid changing a runtime profile shared by other agents.
Now test ZeroClaw before adding Telegram:
zeroclaw agents list
zeroclaw agent -a guide -m "What is 2 + 2? Answer in one sentence."

Checkpoint: ZeroClaw returns a real answer through Ollama. Fix local inference before troubleshooting Telegram.
Step 3: Create your Telegram bot#
Open the verified @BotFather, send /newbot, and choose a display name and a unique username ending in bot. BotFather returns a token and a link to your bot. Use your own new bot; the test bot pictured below is only an example.
Keep the token private. Open the bot link and tap START. At this point, the bot may be silent because its ZeroClaw listener is not running yet.

Checkpoint: your bot’s own chat opens. BotFather creates the bot; ZeroClaw supplies its responses.
Step 4: Connect Telegram to the Ollama agent#
Store the token through ZeroClaw’s masked prompt, enable the channel, and attach it to the agent:
zeroclaw config set channels.telegram.home.bot_token
zeroclaw config set channels.telegram.home.enabled true
zeroclaw config set agents.guide.channels '["telegram.home"]'
Paste the token only into the masked prompt, not into the command itself. The channel-list command above is for this fresh setup; it replaces the list. To preserve existing channels, run zeroclaw config set agents.guide.channels without a value and add telegram.home in the list editor.
For reproducibility, the non-secret settings used in the test are shown below. This is an excerpt, not a replacement for the full Quickstart-generated configuration:
schema_version = 3
[providers.models.ollama.default]
model = "qwen3:0.6b"
uri = "http://127.0.0.1:11434"
think = false
num_ctx = 4096
[agents.guide]
model_provider = "ollama.default"
risk_profile = "locked_down"
runtime_profile = "unbounded"
channels = ["telegram.home"]
[runtime_profiles.unbounded]
agentic = false
[channels.telegram.home]
enabled = true
# Store bot_token using the masked prompt, never in a published example.
Keep the generated locked_down risk profile and other settings. The isolated test supplied the token through the supported process environment override ZEROCLAW_channels__telegram__home__bot_token; the masked prompt above is the persistent setup path documented by ZeroClaw. Neither method requires putting a token in this article or a screenshot.

Checkpoint: agents.guide.model_provider is ollama.default, its channels include telegram.home, and the Ollama URI is local. A bot’s own claim about its model is not proof of the configuration.
Step 5: Pair your account and get a successful reply#
Start one foreground channel listener and leave the terminal open:
zeroclaw channel start
With no authorized peers, this prints a one-time pairing code. In your bot chat, send /bind followed by your own code. The bot should confirm that the account was bound successfully. Then send:
What is 2 + 2? Answer in one sentence.

Observed result on 13 October 2026: the bot replied “2 + 2 = 4.” This confirms the Telegram → ZeroClaw → Ollama → Telegram route for the configuration in Step 4.
Successful pairing saves your numeric Telegram identity in an alias-scoped peer group. In this version, authorization uses peer groups, not the old allowed_users field. Do not use a wildcard peer to get past pairing. Refer to the release-specific Telegram guide for additional users and restart behavior.
Telegram uses outbound long polling here. You do not need a tunnel, a public webhook URL, or an exposed Ollama port. After configuration edits, restart the foreground listener. For long-running use, stop it before switching to zeroclaw daemon or an installed ZeroClaw service, so only one process polls this bot.
If the bot does not reply#
| Symptom | What to check |
|---|---|
| Ollama connection refused | Start Ollama and confirm it is listening on 127.0.0.1:11434 on the same host as ZeroClaw. |
| Model not found | Run ollama list; the configured model must match the downloaded tag exactly. |
| Local chat works, but Telegram stays silent | Check the bot token, channel enabled setting, agent channel binding, and the foreground listener output. |
| Operator approval or pairing message | Redeem the code from the running listener in your own bot chat; check that pairing reports success. |
| Polling conflict | Stop any other listener using the same bot token. |
| Reply is slow or poor | Try a smaller model for connectivity, or a more capable model for quality. Check memory pressure; there is no universal response-time guarantee. |
| Old config keys fail | Confirm zeroclaw --version. This walkthrough uses the v0.8.5 schema and fresh Quickstart aliases. |
For a channel-focused reference, see connecting Telegram to ZeroClaw. For model background, see ZeroClaw with Ollama and Qwen3; compare any older commands there with the tested configuration above.
What this test establishes#
The local model, ZeroClaw agent, Telegram pairing, and reply path worked together with the versions recorded above. The test does not establish production reliability, advanced tool performance, or that Telegram conversations stay on your device. Local inference and Telegram cloud messaging remain two separate parts of this setup.
