modbender Manifest — LLM Router & Observability for OpenClaw
Manifest is an OpenClaw plugin that:
- Routes every request to the most cost-effective model via a 23-dimension scoring algorithm (<2ms latency)
- Tracks costs and tokens in a real-time dashboard
- Sets limits with email alerts and hard spending caps
Source: github.com/mnfst/manifest — MIT licensed. Homepage: manifest.build
Setup (Cloud — default)
Three commands, no coding:
openclaw plugins install manifest
openclaw config set plugins.entries.manifest.config.apiKey "mnfst_YOUR_KEY"
openclaw gateway restart
Get the API key at app.manifest.build → create an account → create an agent → copy the mnfst_* key.
After restart, the plugin auto-configures:
- Registers
manifest/autoas the default model - Injects the
manifestprovider into~/.openclaw/openclaw.json - Starts exporting OTLP telemetry to
app.manifest.build - Exposes three agent tools:
manifest_usage,manifest_costs,manifest_health
Dashboard at app.manifest.build. Telemetry arrives within 10-30 seconds (batched OTLP export).
Verify connection
openclaw manifest
Shows: mode, endpoint reachability, auth validity, agent name.
Setup (Local — offline alternative)
Use local mode only when data must never leave the machine.
openclaw plugins install manifest
openclaw config set plugins.entries.manifest.config.mode local
openclaw gateway restart
Dashboard opens at http://127.0.0.1:2099. Data stored locally in ~/.openclaw/manifest/manifest.db. No account or API key needed.
To expose over Tailscale (requires Tailscale on both devices, only accessible within your Tailnet): tailscale serve --bg 2099
What Manifest Answers
Manifest answers these questions about your OpenClaw agents — via the dashboard or directly in-conversation via agent tools:
Spending & budget
- How much have I spent today / this week / this month?
- What's my cost breakdown by model?
- Which model consumes the biggest share of my budget?
- Am I approaching my spending limit?
Token consumption
- How many tokens has my agent used (input vs. output)?
- What's my token trend compared to the previous period?
- How much cache am I reading vs. writing?
Activity & performance
- How many LLM calls has my agent made?
- How long do LLM calls take (latency)?
- Are there errors or rate limits occurring? What are the error messages?
- Which skills/tools are running and how often?
Routing intelligence
- What routing tier (simple/standard/complex/reasoning) was each request assigned?
- Why was a specific tier chosen?
- What model pricing is available across all providers?
Connectivity
- Is Manifest connected and healthy?
- Is telemetry flowing correctly?
Agent Tools
Three tools are available to the agent in-conversation:
| Tool | Trigger phrases | What it returns |
|---|---|---|
manifest_usage |
"how many tokens", "token usage", "consumption" | Total, input, output, cache-read tokens + action count for today/week/month |
manifest_costs |
"how much spent", "costs", "money burned" | Cost breakdown by model in USD for today/week/month |
manifest_health |
"is monitoring working", "connectivity test" | Endpoint reachable, auth valid, agent name, status |
Each accepts a period parameter: "today", "week", or "month".
All three tools are read-only — they query the agent's own usage data and never send message content.
LLM Routing
When the model is set to manifest/auto, the router scores each conversation across 23 dimensions and assigns one of 4 tiers:
| Tier | Use case | Examples |
|---|---|---|
| Simple | Greetings, confirmations, short lookups | "hi", "yes", "what time is it" |
| Standard | General tasks, balanced quality/cost | "summarize this", "write a test" |
| Complex | Multi-step reasoning, nuanced analysis | "compare these architectures", "debug this stack trace" |
| Reasoning | Formal logic, proofs, critical planning | "prove this theorem", "design a migration strategy" |
Each tier maps to a model. Default models are auto-assigned per provider, but overridable in the dashboard under Routing.
Short-circuit rules:
- Messages <50 chars with no tools → Simple
- Formal logic keywords → Reasoning
- Tools present → floor at Standard
- Context >50k tokens → floor at Complex
Dashboard Pages
| Page | What it shows |
|---|---|
| Workspace | All connected agents as cards with sparkline activity charts |
| Overview | Per-agent cost, tokens, messages with trend badges and time-series charts |
| Messages | Full paginated message log with filters (status, model, cost range) |
| Routing | 4-tier model config, provider connections, enable/disable routing |
| Limits | Email alerts and hard spending caps (tokens or cost, per hour/day/week/month) |
| Settings | Agent rename, delete, OTLP key management |
| Model Prices | Sortable table of 300+ model prices across all providers |
Supported Providers
Anthropic, OpenAI, Google Gemini, DeepSeek, xAI, Mistral AI, Qwen, MiniMax, Kimi, Amazon Nova, Z.ai, OpenRouter, Ollama. 300+ models total.
Uninstall
openclaw plugins uninstall manifest
openclaw gateway restart
This removes the plugin, provider config, and auth profiles. Set a new default model after uninstalling.
Troubleshooting
Telemetry not appearing: The gateway batches OTLP data every 10-30 seconds. Wait, then check openclaw manifest for connection status.
Auth errors in cloud mode: Verify the API key starts with mnfst_ and matches the key in the dashboard under Settings → Agent setup.
Port conflict in local mode: If port 2099 is busy, the plugin checks if the existing process is Manifest and reuses it. To change the port: openclaw config set plugins.entries.manifest.config.port <PORT>.
Plugin conflicts: Manifest conflicts with the built-in diagnostics-otel plugin. Disable it before enabling Manifest.
After backend restart: Always restart the gateway too (openclaw gateway restart) — the OTLP pipeline doesn't auto-reconnect.
Privacy
OTLP telemetry (sent to endpoint):
Fields collected per LLM call: session key, agent name, model name, provider name, token counts (input, output, cache-read, cache-write), tool names, tool success/failure, tool duration, error messages (truncated to classification, no content), message channel, and service metadata. Not collected: user prompts, assistant responses, tool input/output, or any message content.
Routing caveat — manifest/auto sends message content:
When the model is set to manifest/auto, the last 10 non-system messages (including their content) are sent to POST /api/v1/routing/resolve for complexity scoring. This is a separate REST call used only for tier assignment — it is not part of OTLP telemetry. To avoid sending content, disable routing in the dashboard and use a fixed model instead.
Local mode: All data stays on your machine. No external calls are made.
Product analytics: Anonymous usage stats sent to PostHog (hashed machine ID only, no PII). Opt out: MANIFEST_TELEMETRY_OPTOUT=1 or "telemetryOptOut": true in ~/.openclaw/manifest/config.json.