MCP Server

DSAgt exposes its capabilities through a single MCP server, dsagt-server, configured in the per-agent runtime file (.mcp.json for Claude Code, goose.yaml for Goose, and so on) and launched by the agent when it starts. It combines four capabilities (a code registry, a knowledge base, explicit memory, and skill discovery) behind one process with one shared embedder and one ChromaDB.

The 19 tools split across four concerns, all on the one process.

The server's environment

The per-agent MCP config carries an env block with two kinds of variable: routing (the project name and directory, the trace store URI, the embedding backend) and the launching shell's activated environment (PATH, VIRTUAL_ENV, CONDA_PREFIX, PYTHONPATH, the library paths, the module variables, plus any names listed under mcp.env_passthrough in .dsagt/config.yaml). Codex and Cline start the server from that block alone, so it is what makes the server's Python the one with the user's packages. The block is written at dsagt init and dsagt start; after changing the activated environment, run dsagt start (or re-init) so the block matches. A credential never enters the block: a name ending in _KEY, _TOKEN, or _SECRET, or containing SECRET or PASSW, is refused, and dsagt's own service credentials come from the shell or ~/.config/dsagt/env.

Registry tools (5)

Code registration, the readiness reports on record, and pipeline reconstruction. See Provenance for how registered codes are captured.

Tool	Description
search_registry	Semantic search over registered + built-in code specs
get_registry	List every registered code with its MCP-compatible schema
save_code_spec	Register a CLI code as skills/<name>/SKILL.md (executable wrapped with dsagt-run + uv run --with), linked into the agent's native skills directory
reconstruct_pipeline	Render trace_archive/ as a bash script (in run order, output directories created, recorded stdout files redirected) or a Snakemake workflow; output saves it under the project
readiness_reports	The AI-readiness reports on record for a file, each with whether the file is unchanged since that run

Codes are markdown files with YAML frontmatter under <project>/skills/, beside the instruction skills; the executable in the frontmatter is what makes a skill a code. Executables are wrapped with dsagt-run for provenance and uv run --with for Python dependencies.

Knowledge tools (6)

Semantic search and ingestion over indexed document collections. See the Knowledge Base for the retrieval model.

Tool	Description
kb_search	Hybrid semantic search across one or more collections; where filters on a collection's metadata keys, regex and contains on the chunk text
kb_ingest	Index a folder as a new collection; returns a job_id to poll with kb_job_status
kb_append	Add documents to an existing collection (background job)
kb_list_collections	Every collection with its purpose, the metadata keys its chunks carry, and its chunk count
kb_job_status	Check the status of a background ingest/append job
kb_delete_collection	Delete a collection and everything indexed in it; dsagt's own collections are refused

Memory tools (2)

User-confirmed facts that persist across sessions. See Memory.

Tool	Description
kb_remember	Store a user-confirmed fact as an explicit memory (supersedes to replace an outdated one)
kb_get_memories	Retrieve active explicit memories for this project

Skill tools (6)

Discover, install, and author agent skills, and manage external skill sources. See Skills.

Tool	Description
search_skills	Search the external corpus (each hit tagged [catalog · install_skill to add]); skill_name looks up one installed skill
install_skill	Copy a skill from the corpus into <project>/skills/ and link it into the agent's native skills directory
save_skill	Register an agent-authored skill into <project>/skills/<name>/SKILL.md; its scripts become codes, with the specs its codes list declares
add_skill_source	Fetch and index an external skill source (a known name or a Git URL) into the searchable corpus
list_skill_sources	List known and synced external skill sources
delete_skill	Remove an installed skill or registered code: its directory, its native mirror entry, and a code's knowledge-base entry