What it does
banditkeep is a client for a BanditDB server, a contextual bandit database that learns, per campaign, which action tends to succeed in which context. An agent asks for an arm to pick given a context, acts on the answer, and reports a reward between 0 and 1. Other agents using the same server benefit from that outcome immediately, with no retraining step. Campaigns can use LinUCB, which is tunable through alpha, or Thompson Sampling, and can carry budgets for actions with limited cost.
It suits cases where several agents make repeated choices among a set of actions and should share what they learn. The tools cover creating and listing campaigns, inspecting per-arm learning state and convergence, requesting single or batched decisions, recording outcomes, archiving and restoring campaigns, checking budget spend, and adding or pausing actions.
Running it requires a reachable BanditDB server. The base URL defaults to localhost port 8080, and an API key is sent as a bearer token only when one is set. The per-request timeout defaults to 15000 ms. The package offers a CLI, an MCP server and a Docker setup.
Limitations and requirements
Requires a running BanditDB server; the default address is localhost port 8080. Rewards must be between 0 and 1. A batch decision call accepts at most 100 predictions. The HTTP request paths are reconstructions, since the original project does not publish an HTTP API, so they may not match a given server. An optional API key is needed only when the server runs with authentication.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.CloudCorpRecords/banditkeep
Decision memory for agents: bandit campaigns that learn which action works in which context.
v0.1.0As shown on registry.modelcontextprotocol.io when we last read it