Skip to main content

​
Get started in 3 steps

Manifest installs as an OpenClaw plugin and starts tracking costs immediately. Choose cloud mode for hosted dashboards or local mode to keep everything on your machine.
1

Install the plugin

Install Manifest via the OpenClaw CLI:
openclaw plugins install manifest
This downloads the plugin package from npm and registers it with your OpenClaw gateway. The plugin includes:
  • OpenTelemetry SDK for traces, metrics, and logs
  • Smart routing engine with 23-dimension scoring
  • Embedded server for local mode (optional)
  • Agent self-query tools (manifest_usage, manifest_costs, manifest_health)
The plugin auto-detects if you’re using the built-in diagnostics-otel plugin. If both are enabled, you’ll see an error. Disable diagnostics-otel first:
openclaw plugins disable diagnostics-otel
openclaw gateway restart
2

Choose your mode

Manifest runs in two modes: cloud (default) or local.
Cloud mode sends telemetry metadata to the hosted dashboard at app.manifest.build. Your message content never leaves your machine — only OTLP metadata (model, tokens, latency) is sent.Sign up and configure:
  1. Create an account at app.manifest.build
  2. Copy your API key from the dashboard (starts with mnfst_)
  3. Configure the plugin:
openclaw config set plugins.entries.manifest.config.apiKey "mnfst_YOUR_KEY"
openclaw gateway restart
What you get:
  • Multi-device access to your dashboard
  • Multi-agent support with team collaboration
  • Centralized cost tracking across agents
  • No local server management
Keep your API key secure. If compromised, rotate it immediately in the dashboard settings.
3

View your dashboard

After restarting the gateway, telemetry starts flowing immediately.Open the dashboard:
Visit app.manifest.build and log in. You’ll see:
  • Overview — Total tokens, costs, messages, and agents
  • Token usage — Input, output, and cached tokens over time
  • Cost tracking — Spend per model with trend charts
  • Message log — Full conversation history with metadata
  • Agents — Multi-agent dashboard with per-agent sparklines
Telemetry pipeline:
OpenClaw Agent → Manifest Plugin → OTLP Exporter → Backend → Dashboard
Every agent request flows through this pipeline:
  1. Message received — Plugin creates OpenTelemetry trace span
  2. Routing — If model is manifest/auto, scoring algorithm selects tier
  3. Execution — Agent calls the selected model
  4. Completion — Plugin calculates costs, attaches metadata
  5. Export — OTLP exporter batches telemetry and sends to backend (every 5s for traces, 30s for metrics)
  6. Dashboard update — Backend ingests data and updates real-time charts
The gateway batches OTLP exports, so new messages may take 5-30 seconds to appear in the dashboard.

​
Verify the connection

Confirm that Manifest is connected and tracking your agent:

​
Check gateway logs

When the gateway starts, you should see: 
[manifest] Observability pipeline active
[manifest]   Endpoint=https://app.manifest.build/otlp  # or http://127.0.0.1:2099/otlp
[manifest] Connection verified (agent: your-agent-name)
[manifest] Registered as OpenAI-compatible provider (proxy mode)
If you see errors:
  • Cloud mode: Verify your API key starts with mnfst_ and is correctly set
  • Local mode: Check that port 2099 is not in use by another process

​
Use the agent tool

Manifest registers three agent tools for self-query: 
# In your agent chat:
Can you check if Manifest observability is connected?

# Agent response:
{
  "endpointReachable": true,
  "authValid": true,
  "agentName": "my-agent",
  "status": "ok"
}
These tools let your agent query its own usage and costs directly.

​
Enable smart routing

By default, Manifest only observes your requests — it doesn’t change routing. To enable smart routing:

​
Use the manifest/auto model

Set your agent’s model to manifest/auto: 
openclaw config set agents.defaults.model manifest/auto
openclaw gateway restart
Now every request is analyzed and routed to the optimal model automatically.

​
How routing works

When the agent calls manifest/auto:
1

Conversation analysis

The plugin extracts the last 10 messages from the conversation history, excluding system and developer messages (which inflate scores).
2

Scoring across 23 dimensions

The scoring algorithm evaluates:
  • Message count — More messages = higher complexity
  • Conversation depth — How many turns back-and-forth?
  • Token count — Longer prompts = more context
  • Text complexity — Code blocks, technical jargon, math notation
  • Tool usage — Function calling adds complexity
  • Momentum — What tier did the last 5 requests use?
3

Tier assignment

Based on the score, the request is assigned a tier:
  • Fast (score 0-30) — Simple queries, factual questions, summarization
  • Balanced (score 31-60) — Moderate complexity, multi-turn conversations
  • Premium (score 61-100) — Complex reasoning, code generation, long context
4

Model selection

Manifest picks the cheapest available model in that tier from your configured providers. If no provider is set for that tier, it falls back to the next tier up.
Routing adds under 2ms latency to each request. The scoring algorithm is optimized for speed and runs in-memory.

​
Next steps

Now that Manifest is running, explore these features:

Installation guide

Detailed installation instructions for advanced setups, dev mode, and troubleshooting.

Dashboard tour

Learn how to read the analytics, filter by time range, and interpret cost breakdowns.

Configure routing

Set up provider keys, customize tier assignments, and fine-tune routing behavior.

Multi-agent setup

Track multiple agents with separate API keys and cost allocation.

​
Troubleshooting

​
Plugin not loading

Symptom: Gateway starts but no [manifest] logs appear. Fix:
  1. Check plugin is enabled: 
    openclaw plugins list
  2. Verify installation: 
    openclaw plugins install manifest
openclaw gateway restart

​
Connection check failed

Symptom: [manifest] Connection check failed: [error message] Cloud mode fixes:
  • Verify API key is correct and starts with mnfst_
  • Check network connectivity to app.manifest.build
  • Ensure firewall allows HTTPS traffic
Local mode fixes:
  • Check port 2099 is not in use: lsof -i :2099
  • Try a different port: 
    openclaw config set plugins.entries.manifest.config.port 2100
openclaw gateway restart

​
No telemetry in dashboard

Symptom: Dashboard is empty even after sending requests. Fix:
  1. Send a test message to your agent
  2. Wait 30 seconds for OTLP batch export
  3. Check gateway logs for export errors
  4. Verify agent is using a configured model (not manifest/auto without provider setup)

​
Routing not working

Symptom: All requests go to the same model despite using manifest/auto. Fix:
  1. Verify model is set to manifest/auto: 
    openclaw config get agents.defaults.model
  2. Check provider configuration in dashboard → Routing tab
  3. Ensure at least one provider key is configured for each tier
  4. Review gateway logs for routing resolution messages

Need more help?

Join the Discord community for support, bug reports, and feature requests.

Build docs developers (and LLMs) love

Get started for freeTalk to us