album-mcp is the layer that makes scientific software usable by AI agents. Instead of a researcher installing tools, reading docs, and wiring pipelines by hand, an agent does the work: it finds what exists, runs it, chains the results, and writes new tools when none fit. The open protocol is free. We sell the managed version: hosted catalogs, enterprise security, and album-mcp as a managed service.
REST APIs were designed for humans who could read docs, hand-write requests, and parse the output. Agents need something else: structured tool definitions with typed arguments, machine-discoverable schemas, and a uniform invocation contract. MCP (Anthropic's Model Context Protocol) gives us that. album-mcp implements it for the entire scientific-software lifecycle, so any agent that speaks MCP can drive a research pipeline as naturally as it writes a paragraph.
Below is the full surface, organised by category. Each tool is a verb. Operations marked 2FA require a TOTP code; this is how an enterprise gives an agent autonomy with a brake.
An agent receives a single instruction: denoise and segment this microscopy folder. The protocol takes care of everything below.
▸ AGENT denoise & segment /data/batch_042/ ─────────────────────────────────────────────────────────────── ┌─ search_solutions("denoise") │ ✓ matched: n2v-denoising@1.0.2 └─ get_solution_info("n2v-denoising") · args: --input (dir) --output (dir) · env: conda <python=3.10, n2v=0.3.4> ┌─ install_solution("n2v-denoising") [2FA] │ ✓ conda env provisioned (cached) └─ run_solution("n2v-denoising", /data/batch_042/) [2FA] ✓ 128 images denoised → /cache/album/.../denoised/ ┌─ search_solutions("segment") │ ✓ matched: cellpose-segmentation@2.1.0 └─ get_cache_path() // chain inputs ┌─ install_solution("cellpose-segmentation") [2FA] │ ✓ cuda 12 env provisioned └─ run_solution("cellpose-segmentation", ↑denoised) [2FA] ✓ 128 segmentations → /results/segmented/ ─────────────────────────────────────────────────────────────── ▸ DONE elapsed: 4m 12s · no human intervention
album-mcp is shipped as a Docker container that speaks the MCP wire protocol over stdio. Below: clients we have tested live and what they orchestrate well.
| # | Client | Strengths observed | Status |
|---|---|---|---|
| 1 | Claude Desktop | Long-horizon planning. Authoring entire solution files. Validating before deploy. | verified |
| 2 | Cursor | Inline scaffolding while editing solution code. RDF lint feedback in editor. | verified |
| 3 | Windsurf | Multi-agent orchestration across catalogs. Chaining tools by cache path. | verified |
| 4 | Custom MCP | Bring your own. The wire format is open. | supported |
An autonomous agent can do real damage at the speed of a tokeniser. We gate the dangerous operations behind a TOTP. The agent plans freely; it acts only when an authenticated human (or service-account secret) is present. The granularity is per-tool, per-session.
free / discoverable gated / 2FA ────────────────── ──────────── list_catalogs ┊ install_solution list_solutions ┊ run_solution search_solutions ┊ uninstall_solution get_solution_info ┊ deploy_solution get_cache_path ┊ remove_solution get_solution_scaffold ┊ assemble_solution ┊ 11 free · 5 gated validate_solution ┊ add_catalog ┊ remove_catalog ┊ update_catalog ┊
An agent that can install is interesting. An agent that can publish is dangerous. An agent that can publish behind 2FA is shippable. S. Proft & J. P. Albrecht, design memo, 2026
An interactive transcript. Type a prompt; watch album-mcp dispatch.