What it does
Phases Agents is a local MCP server that profiles a project, validates a catalogue of skills against a strict contract, and returns a selection plan. Given the same target, catalogue and parameters, it returns the same decision. The server only selects and exposes skills; the calling model reads them and acts with its own tools. It never executes a skill. It exposes five tools, and get_skill takes a skill identifier rather than a path. Absolute paths and detected secrets are masked in public output, and encoded responses stay under 1 MiB.
It suits work where the same question must produce the same plan, such as audit and compliance tasks, rather than open-ended brainstorming. Each skill is a package with a phases.json manifest and a SKILL.md file containing fourteen required sections. Skills in legal, compliance or regulatory domains must cite an official source, a jurisdiction and a verification date for every rule they reference.
The server uses only the Python standard library, with no runtime dependencies. It can be installed as a command or started directly with Python. A skills-roots.json file must point to the directory holding the skill packages. The server reads JSON-RPC messages from standard input, and MCP clients such as Claude Code connect through a .mcp.json file. No token or environment variable is needed. The B3 plan format is recommended for new work, and the legacy flat format is still supported.
Limitations and requirements
A skills-roots.json file pointing to a skills directory is required. Python is required to run the server. The legacy plan format (without plan_version) will be deprecated before removal. client_capabilities is accepted only in the B3 format.
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
Phases Agents
Deterministic skill selection: local discovery, strict validation, every skill classified.
v0.5.0As shown on registry.modelcontextprotocol.io when we last read it