Banditkeep MCP Server

banditkeep

Give a swarm of agents a memory for decisions. banditkeep is a small client for a BanditDB server: a contextual bandit database that learns, per campaign, which action tends to pay off in which context. An agent asks which arm to pick, acts on the answer, then reports whether it worked — and every later agent pointing at the same server benefits from that outcome immediately, with no retraining step.

Two exploration algorithms anchor the campaigns: LinUCB, deterministic and tunable through alpha, and Thompson Sampling, which explores naturally in a Bayesian way and needs no tuning. Campaigns can also carry budgets for actions that cost something limited, so spend is steered across a horizon instead of running dry on the first day.

Configuration

Three environment variables control the client:

• BANDITDB_URL (optional) — base URL of your BanditDB server. Defaults to http://localhost:8080.
• BANDITDB_API_KEY (optional) — when set, it is sent on every request as Authorization: Bearer <key>. Omit it when your server runs without auth; leaving it unset is never an error by itself.
• BANDITDB_TIMEOUT_MS (optional) — per-request timeout in milliseconds, enforced with an AbortController. Defaults to 15000. (This variable is this rebuild's own addition.)

Tools

• create_campaign — create a new decision campaign. Requires campaign_id and arms; optionally takes features (names, in order) or feature_dim, alpha, algorithm ("linucb" or "thompson_sampling"), metadata, budgets, and decay_half_life_hours.
• list_campaigns — list all active campaigns, showing algorithm and alpha; useful to check what exists before calling get_intuition.
• campaign_diagnostics — inspect per-arm learning state: theta norms, prediction counts, reward rates, and entropy health. Reach for it when a campaign does not seem to be learning or one arm is dominating.
• campaign_report — a business-level convergence report: whether the campaign has statistically converged, and which arm is winning, with confidence intervals.
• get_intuition — ask which arm to pick for a given context (an object of feature-name to number, or an array of numbers). Optional eligible_arms / exclude_arms narrow the choice for this request only. Returns the arm and an interaction_id to save.
• batch_get_intuition — get decisions for multiple campaigns in a single round-trip: pass predictions as a list of {campaign_id, context} dicts, at most 100 per call.
• record_outcome — report whether the chosen action succeeded (1.0) or failed (0.0), closing the loop for an interaction_id. Rewards outside [0, 1] are rejected locally with reward must be between 0 and 1.
• archive_campaign — soft-delete a campaign: it pauses predictions and rewards but preserves all learned weights.
• restore_campaign — restore an archived campaign to active status with all weights intact.
• campaign_budget — budget spend per resource: consumed, remaining, utilization, shadow price, and exhaustion.
• add_arm — add an action to a live campaign, optionally warm-started from existing arms and with budget costs.
• set_arm_status — pause, retire, or reactivate an action (status is "active", "paused", or "retired"); the arm keeps everything it learned.

Unset optional arguments are omitted from requests entirely. Responses come back exactly as the server returned them. An upstream non-2xx reply fails with banditdb error: HTTP <status> and an unreachable server (including a timeout) with banditdb unreachable.

Endpoint provenance

The original project's README and MCP documentation pin the twelve tools, their arguments, the configuration above, and the decision-loop semantics — but no HTTP API is published in those sources. The request paths below are reconstructions, labelled as such:

• POST <base>/campaigns with the arguments as the JSON body (used by create_campaign).
• GET <base>/campaigns (used by list_campaigns).
• GET <base>/campaigns/{campaign_id}/diagnostics (used by campaign_diagnostics).
• GET <base>/campaigns/{campaign_id}/report (used by campaign_report).
• GET <base>/campaigns/{campaign_id}/pacing (used by campaign_budget; the SDK method is named pacing).
• POST <base>/predict with body {"campaign_id": ..., "context": ...} (used by get_intuition).
• POST <base>/predict/batch with body {"predictions": [...]} (used by batch_get_intuition).
• POST <base>/reward with body {"interaction_id": ..., "reward": ...} (used by record_outcome).
• POST <base>/campaigns/{campaign_id}/archive with body {} (used by archive_campaign).
• POST <base>/campaigns/{campaign_id}/restore with body {} (used by restore_campaign).
• POST <base>/campaigns/{campaign_id}/arms with body {"arm_id": ...} (used by add_arm).
• POST <base>/campaigns/{campaign_id}/arms/{arm_id}/status with body {"status": ...} (used by set_arm_status).

Path segments are sent verbatim. One discrepancy worth knowing: the vendor's MCP docs page lists nine of the twelve tools (it omits campaign_budget, add_arm, and set_arm_status); the package's own MCP table enumerates all twelve, and this rebuild follows the package table.

CLI

Each tool is a subcommand taking one JSON argument string:

npx @reneturcios/banditkeep create_campaign '{"campaign_id":"model-routing","arms":["fast","cheap"],"feature_dim":4}'
npx @reneturcios/banditkeep list_campaigns '{}'
npx @reneturcios/banditkeep campaign_diagnostics '{"campaign_id":"model-routing"}'
npx @reneturcios/banditkeep campaign_report '{"campaign_id":"model-routing"}'
npx @reneturcios/banditkeep get_intuition '{"campaign_id":"model-routing","context":[0.8,0.2,0.5,0.1]}'
npx @reneturcios/banditkeep batch_get_intuition '{"predictions":[{"campaign_id":"model-routing","context":[0.8,0.2,0.5,0.1]}]}'
npx @reneturcios/banditkeep record_outcome '{"interaction_id":"9f2c-4e1a","reward":1.0}'
npx @reneturcios/banditkeep archive_campaign '{"campaign_id":"model-routing"}'
npx @reneturcios/banditkeep restore_campaign '{"campaign_id":"model-routing"}'
npx @reneturcios/banditkeep campaign_budget '{"campaign_id":"model-routing"}'
npx @reneturcios/banditkeep add_arm '{"campaign_id":"model-routing","arm_id":"balanced"}'
npx @reneturcios/banditkeep set_arm_status '{"campaign_id":"model-routing","arm_id":"cheap","status":"paused"}'
npx @reneturcios/banditkeep --version

Output is the tool's result text (the upstream body on success). Exit codes: 0 on success, 1 on a runtime error, 2 on a usage error.

MCP server

npx -y @reneturcios/banditkeep mcp

MCP client config (provide the key through the client's environment when your server uses one):

{
  "mcpServers": {
    "banditkeep": {
      "command": "npx",
      "args": ["-y", "@reneturcios/banditkeep", "mcp"],
      "env": {
        "BANDITDB_URL": "http://localhost:8080",
        "BANDITDB_API_KEY": "<your key>"
      }
    }
  }
}

Docker

docker run -i ghcr.io/cloudcorprecords/banditkeep
docker run --rm ghcr.io/cloudcorprecords/banditkeep list_campaigns '{}'

Development

npm install
npm run build   # tsc -> dist/
npm test        # tsx --test test/*.test.ts (fake fetch only, no network)
npm run dev     # run the CLI from source

License

MIT — Copyright (c) 2026 Rene Turcios