Open source

Sluice: your own API traffic, over MCP

A local tool that records the API calls your own signed-in apps make, rebuilds each service's structure from them, and lets an agent read the result over MCP. Nothing leaves the machine, and every replay passes read-only rails enforced below every caller.

Role
Design & development
Timeline
2026 · open source (AGPL-3.0)
Built with
TypeScript, SQLite, MCP, Chrome DevTools Protocol
Link
View on GitHub
Diagram: browser capture, a local proxy and your own session merge into one local store that an agent reads over MCP.

Give an agent your own data, with hard limits

Much of a person's working context lives in the web apps they use every day. Sluice explores how much an agent can do with the traffic those apps already exchange on your own machine, without a cloud service in between.

The same mechanisms that make this possible could do harm, so the limits had to be part of the design rather than a setting: what may leave the machine, what may be stored, and what a replay is allowed to do.

Capture, normalise, then expose read-only

Four capture modes feed one data model: a Chrome instance driven over the DevTools Protocol, a browser extension limited to hosts you name, a local pass-through proxy that never modifies traffic, and on-demand replay of your own session. Each capture is redacted again before it reaches SQLite.

Nine app adapters normalise the traffic into workspaces, containers, items and edges. A local web app, a CLI and an MCP server read the same store.

  • MCP server: 11 core tools plus 26 contributed by the nine apps, 37 with every app enabled
  • Replays: three HTTP methods, a write and admin denylist, the app's own hosts only, 60 requests per 60 seconds
  • Credentials stay in the memory of the process that read them; the schema has no column for one
  • No telemetry, crash reports or update checks: the only outbound requests are your own

Selected contributions

  • A project graph for coding agents

    A second MCP server gives agents working on the codebase a local knowledge graph of it: packages, files, symbols, imports and calls, routes, tables, tests and reviewed findings. It publishes 10 tools, 3 resources and an orientation prompt.

    Retrieval is deterministic, GraphRAG-style: BM25 over SQLite full-text search picks the seeds, exact and partial name matches are added, candidates are fused and ranked, and a bounded walk over typed relationships supplies the connected context. There is no LLM, embedding API or vector database in the loop, and a content fingerprint tells the agent when the graph is stale.

  • Rails below every caller

    The replay limits sit beneath the UI, the CLI and the MCP tools alike, so no interface can route around them. App-contributed MCP tools only get a read-only projection of the store and never touch the database directly.

Public and tested

Sluice is public on GitHub under AGPL-3.0, with 73 test files and CI on the main branch. It is not published to npm yet; it installs from a clone.

Checkable limits

The project asks to be judged on its mechanisms rather than its promises: no egress path in the codebase, no credentials on disk, and a standing decision never to add detection evasion. Writing those limits down and making them checkable in the source was as much of the work as the capture engines.