Does it send workspace data over the network?

No. The executable Python scripts make no network requests, use local JSON state under the OpenClaw workspace memory directory, and do not use eval, dynamic imports, downloaders, or shell command execution.

How does model routing work?

Routine Anthropic tasks map to anthropic/claude-sonnet-4-5, while only explicitly complex reasoning maps to anthropic/claude-opus-4-5. The bundled model_router.py accepts a task prompt and returns routing guidance.

Will it overwrite my OpenClaw files automatically?

No. AGENTS.md generation prints to stdout unless an output option is supplied, and heartbeat installation previews by default. The explicit install command backs up an existing HEARTBEAT.md before writing.

Documents

OpenClaw Token Optimizer

Reduce OpenClaw token waste with lazy context, model routing, heartbeat planning, and local budget checks.

What it does

Reduce recurring token waste in OpenClaw 2026.6.x by recommending prompt-specific context, routing routine Anthropic work to Sonnet and complex reasoning to Opus, and planning due heartbeats around quiet hours and cache TTL. Local scripts also track daily token budgets, preview file changes, and back up an existing HEARTBEAT.md before replacement.

When to use it

Reducing context in large multi-agent workspaces
Keeping routine tasks on Sonnet
Scheduling heartbeat checks around quiet hours
Checking local daily token budgets

The skill document

OpenClaw Token Optimizer v3.2.0

Version License

Built for current OpenClaw 2026.6.x agents by Asif2BD · GitHub · Security Notes

Security notice: local-only optimization toolkit. The Python scripts make no network requests and do not execute dynamic code. Commands that write files are explicit, documented, and backup-safe.

Install

openclaw skills install @asif2bd/openclaw-token-optimizer

What's New in v3.2

Modern ClawHub Card

The skill card now leads with a clearer product promise, current OpenClaw compatibility, security posture, and practical commands instead of the older long-form v3.0 copy.

Current OpenClaw Model Routing

Routine Anthropic work maps to anthropic/claude-sonnet-4-5; only explicitly complex reasoning routes to anthropic/claude-opus-4-5.

Audit-Safe Writes

context_optimizer.py generate-agents prints to stdout by default.
Writing an optimized AGENTS.md requires --output or --workspace-output.
./scripts/optimize.sh heartbeat previews only by default.
./scripts/optimize.sh heartbeat install writes HEARTBEAT.md only after backing up an existing file.

What It Optimizes

Context Loading

Stop loading every workspace file into every session. Recommend only the context needed for the current prompt.

python3 scripts/context_optimizer.py recommend "hi"

Typical result: SOUL.md + IDENTITY.md only for a greeting.

Model Routing

Keep standard work on Sonnet and reserve Opus for deep reasoning.

python3 scripts/model_router.py "thanks!"
python3 scripts/model_router.py "design a multi-agent architecture"

Heartbeat Scheduling

Run only checks that are due, respect quiet hours, and align heartbeat timing with prompt-cache TTL.

python3 scripts/heartbeat_optimizer.py plan
python3 scripts/heartbeat_optimizer.py cache-ttl

Token Budgets

Track local daily budget state and surface warnings before runaway spend.

python3 scripts/token_tracker.py check

Quick Start

1. Find the minimum context for a request

python3 scripts/context_optimizer.py recommend "write a deployment checklist"

2. Preview an optimized AGENTS.md

python3 scripts/context_optimizer.py generate-agents

Save a review copy only when you explicitly ask for it:

python3 scripts/context_optimizer.py generate-agents --output ~/.openclaw/workspace/AGENTS.md.optimized

3. Preview heartbeat installation

./scripts/optimize.sh heartbeat

Install with backup protection:

./scripts/optimize.sh heartbeat install

4. Route a task

python3 scripts/model_router.py "debug this error"

Savings Profile

Area	Typical waste	Optimizer approach
Startup context	Loading docs, memory, tools, and old logs every turn	Lazy-load only what the prompt needs
Routine model use	Opus used for quick replies or simple checks	Sonnet for routine work, Opus for complex reasoning
Heartbeats	Every check runs on every heartbeat	Due-check planner + quiet hours
Cache TTL	Idle sessions trigger expensive cache rewrites	55-minute heartbeat guidance for 1-hour cache TTL
Budget control	Spend discovered after the fact	Local budget checks before expensive work

Expected savings: 50-80% in context-heavy OpenClaw workspaces, with context optimization usually delivering the biggest win.

Files

Scripts

scripts/context_optimizer.py — prompt classification and context bundle recommendations
scripts/model_router.py — Sonnet/Opus-aware task routing
scripts/heartbeat_optimizer.py — due-check planning, quiet hours, cache-TTL interval guidance
scripts/token_tracker.py — local budget tracking
scripts/optimize.sh — convenience wrapper around bundled scripts

Assets

assets/HEARTBEAT.template.md — backup-safe heartbeat template
assets/cronjob-model-guide.md — scheduled-task model guidance
assets/config-patches.json — native and optional config examples

References

references/PROVIDERS.md — optional external-provider strategy notes
SECURITY.md — complete security and data-handling breakdown

Security Posture

No network calls in executable scripts
No eval, dynamic imports, shell command execution, or downloader behavior
Local JSON state only under the OpenClaw workspace memory directory
Workspace writes require explicit commands
Existing HEARTBEAT.md is backed up before replacement
.clawhubsafe covers every published file with SHA256 hashes

Verify locally:

sha256sum -c .clawhubsafe

Native OpenClaw Diagnostics

Use OpenClaw's built-in context and usage commands alongside this skill:

/context list
/context detail
/usage tokens
/usage cost

Use the built-ins to see where tokens are going, then use Token Optimizer to reduce the recurring waste.

Best Fit

Multi-agent OpenClaw workspaces with large instruction files
Agents that run heartbeats or scheduled checks
Teams trying to keep Opus for hard work instead of routine work
xCloud, MissionDeck, or other hosted OpenClaw deployments where every saved prompt token compounds across users

More by Asif2BD

openclaw skills install @asif2bd/jarvis-mission-control
openclaw skills search asif2bd

Questions people ask

Does it send workspace data over the network?: No. The executable Python scripts make no network requests, use local JSON state under the OpenClaw workspace memory directory, and do not use eval, dynamic imports, downloaders, or shell command execution.
How does model routing work?: Routine Anthropic tasks map to anthropic/claude-sonnet-4-5, while only explicitly complex reasoning maps to anthropic/claude-opus-4-5. The bundled model_router.py accepts a task prompt and returns routing guidance.
Will it overwrite my OpenClaw files automatically?: No. AGENTS.md generation prints to stdout unless an output option is supplied, and heartbeat installation previews by default. The explicit install command backs up an existing HEARTBEAT.md before writing.