Connect Ollama to Observagent.

1. Prepare Python and Ollama

Run these checks on your Windows computer:

python --version
ollama list

If Python is not found, try py --version. If neither command works, install Python, then close and reopen PowerShell:

winget install --exact --id Python.Python.3.13 --source winget

If winget is unavailable, use the Windows installer from python.org. Enable Add Python to PATH if offered. After reopening PowerShell, verify python --version; use py for the initial environment-creation command if only that launcher works.

Open the Ollama application and check that the server is listening:

Invoke-RestMethod "http://127.0.0.1:11434/api/version"

Expect a version number. If the connection fails, start Ollama before continuing. If using ollama serve manually, keep it in a separate terminal without the proxy's OLLAMA_HOST setting.

2. Get the proxy repository and create an environment

Clone the Observagent Ollama proxy repository, or use GitHub's Download ZIP option and extract it. The repository includes the proxy, pinned dependencies, offline tests, delivery verifier, license, and complete setup guide.

Set-Location $HOME
git clone https://github.com/Observagent/observagent-ollama-proxy.git
Set-Location "$HOME\observagent-ollama-proxy"
python -m venv .venv

Run each installation command separately. The last package name is one complete name; do not insert a newline inside it. Calling the environment's Python directly means activation is optional.

.\.venv\Scripts\python.exe -m pip install fastapi uvicorn httpx
.\.venv\Scripts\python.exe -m pip install opentelemetry-sdk
.\.venv\Scripts\python.exe -m pip install opentelemetry-exporter-otlp-proto-http

Alternatively, install the pinned dependencies from the repository file:

.\.venv\Scripts\python.exe -m pip install -r requirements.txt

You may activate with .\.venv\Scripts\Activate.ps1 and use python, but all remaining commands work without activation. If activation is blocked by execution policy, use the direct commands below; no policy change is needed.

3. Create or select an Observagent project

Sign in to your Observagent dashboard. Open Settings / Projects, create an Ollama project or select an existing one, enable Logs, Metrics, and Traces, and turn monitoring on. Choose Connect source, then Generate / replace project key. Copy the full key when shown; it needs telemetry:write permission.

A separate Ollama project can make its telemetry easier to identify and filter.

If you replace a project key, update every sender that used the previous key.

If project-key creation is unavailable, contact your Observagent administrator or support channel. Do not use your dashboard password or a model-provider key.

4. Configure Window 1 and start the proxy

Run these commands in the same PowerShell window that will start the proxy. Enter your real endpoint at the prompt, such as https://dashboard.observagent.cloud/v1/traces for that deployment. Use the full /v1/traces URL, not the marketing site or documentation URL.

Set-Location "$HOME\observagent-ollama-proxy"
$env:OTEL_EXPORTER_OTLP_TRACES_ENDPOINT = Read-Host "Full backend trace URL ending in /v1/traces"
$env:OBSERVAGENT_API_KEY = (Read-Host "Paste the complete raw Observagent project API key").Trim()
$env:OBSERVAGENT_SESSION_ID = "ollama-cli-test"

Supply the complete project key to OBSERVAGENT_API_KEY; a shortened key prefix cannot authenticate requests. Do not include Authorization=Bearer, %20, quotes, or a leading #. The proxy derives the logs and metrics endpoints from the trace endpoint. You can override them with OTEL_EXPORTER_OTLP_LOGS_ENDPOINT and OTEL_EXPORTER_OTLP_METRICS_ENDPOINT. All three exporters use the same project API key.

Check that the settings exist without printing the key:

Write-Host "Endpoint:" $env:OTEL_EXPORTER_OTLP_TRACES_ENDPOINT
Write-Host "API key configured:" (-not [string]::IsNullOrWhiteSpace($env:OBSERVAGENT_API_KEY))

Run the offline tests, then start the proxy:

.\.venv\Scripts\python.exe test_ollama_proxy.py
.\.venv\Scripts\python.exe -m uvicorn ollama_proxy:app --host 127.0.0.1 --port 11435

Expect the API-key check to return True, then Application startup complete and Uvicorn running on http://127.0.0.1:11435. Leave this window open. Settings are local to this PowerShell process and disappear when you close it. Variables set in a different terminal or an unloaded .env file do not configure the proxy.

5. Connect Window 2 and generate activity

Open a second PowerShell window on the same computer:

$env:OLLAMA_HOST = "http://127.0.0.1:11435"
ollama list

Choose an installed model name from the list:

$model = Read-Host "Installed model name from ollama list"
ollama run $model

Send a prompt, for example: Explain telemetry in two sentences. Let the answer finish. OLLAMA_HOST applies to the CLI in this window; keep the real Ollama server on 11434. Requests from other windows that bypass the proxy are not captured.

Other Ollama HTTP clients must also target the proxy at http://127.0.0.1:11435. Use the client's documented Ollama base-URL setting when it does not honor OLLAMA_HOST.

If the client produces a model response but Window 1 shows no POST /api/chat, the client bypassed the proxy and probably connected directly to port 11434.

6. Verify logs, metrics, and traces in Observagent

Switch to Window 1 with Alt+Tab or select its terminal tab. After a completed response, expect ollama.request.completed: input=... output=...; logs, metrics and trace queued for export. Queued means buffered for export, not confirmed backend storage.

In the dashboard, sign in with the account associated with the key. Open Telemetry Data, choose All sources and All available, select Logs or All, clear old filters, and refresh. Search for agent:ollama-cli or event:ollama.request.completed. Inspect a log's original attributes to see the model, input_token_count, output_token_count, and durations.

Check logs, metrics, and traces separately. Metrics include ollama.request.count, ollama.request.duration, and ollama.token.count when usage is available. One signal arriving does not prove that the others were stored.

Log events are ollama.request.completed, ollama.request.failed, and ollama.request.incomplete. Errors before connection produce a failed request log. There are no logs for health checks or model-list requests. Earlier trace-only versions of the proxy do not send logs; replace the script and restart to use this version.

7. Diagnose delivery directly

In Window 1, press Ctrl+C to stop the proxy while keeping its environment settings. Run the supplied diagnostic file to send one synthetic log, metric, and trace with session ollama-delivery-test:

.\.venv\Scripts\python.exe verify_delivery.py

This creates test records. Expect HTTP 200 and count: 1 for each signal. That confirms authenticated delivery for the configured key independently of Ollama. A successful test does not prove that automatic proxy exports are working; restart the proxy and generate another real request to check that too.

To test only routing and authorization without storing records, use the empty-batch mode:

.\.venv\Scripts\python.exe verify_delivery.py --empty

Expect count: 0. HTTP 200 with HTML is not an ingestion confirmation. HTTP 401 means the complete raw key was not accepted; generate a fresh project key and copy it exactly. The prefix alone can never authenticate. HTTP 403 means the key or project is not permitted to send that signal. A 404 usually indicates a wrong path. A timeout or TLS error needs a connectivity, domain, or certificate check. Keep certificate verification enabled.

If the test stores records but they are missing from project history, check the key's project, signed-in account, filters, and whether the website and ingestion URL refer to the same deployment. Do not regenerate keys repeatedly to diagnose a display issue.

8. Change keys, restart, or stop collection

To route new data to a different project, stop the proxy in Window 1 with Ctrl+C, then update the key and restart:

$env:OBSERVAGENT_API_KEY = Read-Host "Paste the new project API key"
.\.venv\Scripts\python.exe -m uvicorn ollama_proxy:app --host 127.0.0.1 --port 11435

To start in a new PowerShell window, repeat the endpoint, key, and session settings from step 4. Do not paste keys into chats, screenshots, committed files, or public documentation.

To bypass the proxy, exit the interactive Ollama session and run this in Window 2 before starting another CLI session:

Remove-Item Env:OLLAMA_HOST -ErrorAction SilentlyContinue

Press Ctrl+C in Window 1 to stop the proxy and flush queued exports during normal shutdown. Stopping collection does not delete already stored records.

Troubleshooting the local setup

• Python was not found: install Python, reopen PowerShell, and verify python or py before creating the environment. Activation cannot work until environment creation succeeds.
• Invalid requirement ending in otlp-: the package name was split. Run the complete opentelemetry-exporter-otlp-proto-http installation line from step 2.
• Set OTEL_EXPORTER_OTLP_TRACES_ENDPOINT and OBSERVAGENT_API_KEY first: one or both variables are missing in the window starting the proxy. Run step 4 again there. A blank endpoint in another terminal tells you nothing about the proxy process.
• Boolean environment checks return false: remove any leading # from the assignment and run it in the same PowerShell window that starts the relevant process. A .env file is not automatically loaded by this proxy.
• Address already in use on 11435: another proxy instance may still be running. Return to its window and stop it before restarting.
• Cannot reach Ollama on 11434: start the Ollama application. Do not start the Ollama server using the CLI window's proxy address.
• Proxy receives no request: configure the client with $env:OLLAMA_HOST = "http://127.0.0.1:11435" or its explicit Ollama base-URL option. A local report proves inference completed, not that the proxy saw it.
• Export returns 401: run verify_delivery.py first. Confirm OBSERVAGENT_API_KEY contains the complete raw key, not the prefix or an Authorization=Bearer%20... header. Revoke any key pasted into source code, logs, screenshots, shell history, or support messages.
• Only some signals arrive: pull the latest proxy version from GitHub, restart it, enable Logs, Metrics, and Traces for the project, and send a new request.
• Telemetry arrives but a dashboard total stays unchanged: confirm that the selected view and filters include recent Ollama events.
• PowerShell remains at the >> prompt after a pasted Python block: press Ctrl+C. A here-string closing marker must start at column one and Python imports must not be split incorrectly. Use verify_delivery.py instead of pasting a multiline Python here-string.

Escalate an unresolved connection problem

If a newly generated key still fails verify_delivery.py, contact your Observagent administrator or support channel. Share only the request timestamp, endpoint hostname, HTTP status, project name, and non-secret key prefix. Never share the complete raw key. Server configuration, account validation, and storage diagnostics belong in the private operator runbook and are intentionally excluded from this public guide.

Files and reference

• ollama_proxy.py: local forwarding service and batched log, metric, and trace exporters.
• requirements.txt: runtime dependencies.
• test_ollama_proxy.py: offline regression checks.
• verify_delivery.py: direct synthetic or empty-batch delivery checks.
• README.md: this setup and troubleshooting guide.

Ollama reports usage in the final streaming response; see Ollama usage documentation. See the Observagent Ollama integration guide for the public walkthrough.