BanditDB

BanditDB is an in-memory decision database for AI agents — real-time learning from outcomes. Use it to auto-tune notification timing, model routing, or promp...

Security Scan

The skill's declared purpose (an in-memory decision DB) matches its instructions (run a local BanditDB service and call its HTTP API); nothing requested by the skill is disproportionate to that purpose.

Purpose & Capability

Name/description describe a local decision database. SKILL.md and references/api.md instruct the agent to create campaigns, call /predict, and post /reward to a locally-hosted service — these requirements align with the stated purpose.

Instruction Scope

Instructions are scoped to interacting with a local HTTP service (default http://localhost:8080) and do not ask the agent to read unrelated files or credentials. They do mention registering as an MCP server (Claude/Cursor/etc.), which implies optional network exposure if the agent is configured to register with external MCP hosts; this is a behavioral expansion worth confirming before enabling that feature.

Install Mechanism

This is an instruction-only skill (no install spec). The docs point users to external artifacts (GitHub releases, Docker image, PyPI). That is normal for this kind of skill, but installing binaries/images from external sources should be validated (see user guidance). Also note a minor provenance mismatch: the GitHub release org is listed as dynamicpricing-ai while the Docker image is simeonlukov — not necessarily malicious but worth verifying.

Credentials

No environment variables, credentials, or config paths are requested by the skill. The API interactions are to localhost by default, so no external secrets access is required by the skill itself.

Persistence & Privilege

always is false and model invocation is allowed (default). The skill does not request permanent platform-level privileges or modifications to other skills; no elevated persistence is requested.

Assessment

This skill appears internally coherent: it expects you to run a local BanditDB service and call its HTTP API. Before installing/using it: 1) Verify the upstream artifacts — prefer official GitHub releases and check release signatures or checksums; confirm the Docker image owner (simeonlukov) is the intended publisher for the project (dynamicpricing-ai vs simeonlukov mismatch). 2) If you run the Docker image or binary, run it with least privilege and restrict network exposure (it binds to port 8080 by default — avoid exposing that port publicly). 3) If you enable MCP registration, only register with trusted MCP hosts because that step can expose interactions beyond localhost. 4) If you plan to install the banditdb-python package, review the package on PyPI for authenticity. These steps will reduce supply-chain and network exposure risk.

BanditDB Skill

BanditDB is a self-hosted decision database. It learns which choice works best for which context through contextual multi-armed bandits — no ML pipeline required.

Setup

Install BanditDB from GitHub releases or run the Docker image (see references/api.md for details). Default port: 8080. Verify by requesting GET /campaigns.

Core Workflow

Three-step loop — create once, then predict and reward repeatedly:

1. Create a campaign — define a campaign ID, the arms (choices), and context feature dimension.
2. Get a prediction — pass a context vector, receive the recommended arm and an interaction ID.
3. Record a reward — report the outcome (0.0–1.0) for the interaction ID.

For full API details, request/response examples, and MCP tool registration, see references/api.md.

Designing Context Vectors

The context vector is the most important design decision. Each float encodes something about the current situation. Normalize values to roughly 0–1 range.

Examples:

• Notification timing: [hour_of_day/24, day_of_week/7, messages_today/10, last_response_delay_mins/60]
• Tool selection: [query_length/500, has_code_mention, has_url, specificity_score]
• Prompt strategy: [task_complexity, domain_familiarity, output_length_needed, structured_output]

Use Cases for OpenClaw Agents

• Smart notifications — learn when/how to reach the user (arms: morning/afternoon/evening, channel variants)
• Tool routing — which tool to use for a query type (arms: web_search/memory/file_lookup/ask)
• Model selection — which model for which task (arms: opus/sonnet/haiku)
• Response style — learn user preferences (arms: brief/detailed/bullet_points)
• Heartbeat frequency — when to check in vs stay quiet

Key Details

• Algorithms: LinUCB (default, supports causal analysis) or Thompson Sampling
• Cold start: meaningful lift typically after 300–1500 interactions depending on noise
• Parquet export available for offline causal analysis (LinUCB only)
• WAL ensures crash recovery — no data loss on restart
• ~10K predictions/s on a single node