Skip to content

Agent Integration

Jean-Baptiste THERY edited this page Jul 15, 2026 · 30 revisions

Agent integration

Ragmir is a confidential local RAG for coding agents and scripts that need cited retrieval from project documents through the CLI, TypeScript, or a stdio MCP server. Core indexes selected files on the workstation, keeps the corpus and generated index there, works offline by default with local-hash, and does not upload source documents or call a model.

Choose the handoff

Path What stays local What crosses the boundary
Hosted coding agent Corpus, index, and retrieval Only passages returned to the consumer
Local coding agent or script Corpus, index, retrieval, and consumer Nothing unless that consumer uses a network service
Ragmir Chat Corpus, index, retrieval, and answer generation One explicit model download during setup, then no network

Install native helpers

rgr setup --agents claude,codex,kimi,opencode,cline
rgr doctor

Setup creates an ignored .ragmir/run.cjs project runner, local MCP helpers, and project-scoped skill links for the selected clients. Re-run installation when you need a different scope or copy mode:

rgr install-agent --agents codex,claude

Project scope is the default. --mode copy is a fallback for filesystems that cannot follow symlinks. Ragmir refuses to overwrite unmanaged same-name skills unless --force is explicitly selected.

MCP workflow

Read ragmir://context first when the client supports resources. It identifies the active knowledge base, readiness, freshness, source coverage, and available operations.

Resource Use
ragmir://context Active base identity, readiness, freshness, coverage, and capabilities
ragmir://sources Source coverage, skipped-file counts, and index drift, with complete totals and per-file lists capped at 50

Use compact retrieval first, then expand one returned citation only when more context is required:

  1. ragmir_status verifies the active base.
  2. ragmir_route_prompt decides whether retrieval is useful.
  3. ragmir_search, ragmir_ask, or ragmir_research returns bounded cited evidence.
  4. ragmir_expand opens one exact indexed chunk with a small neighbor window.
  5. ragmir_audit and ragmir_security_audit verify coverage and privacy posture.

The server also exposes ragmir_evaluate for local golden-query checks and ragmir_usage_report for metadata-only retrieval usage. Search, ask, research, expansion, audit, and evaluation accept an optional maxBytes. Every tool and resource JSON response stays below the configured mcpMaxOutputBytes budget and an absolute 1 MiB ceiling.

ragmir_ask returns cited evidence, not a model-generated answer. A hosted consumer receives the passages it requests, so use a local consumer when that handoff is not acceptable.

Every tool advertises non-destructive behavior. Search, ask, research, and evaluation conservatively advertise open-world behavior because explicitly enabled semantic models may download public weights. The prompt router, security audit, and usage report also advertise read-only, idempotent behavior. Other tools conservatively do not because they can initialize ignored local state or append metadata-only access logs. ragmir_evaluate accepts only an existing project-relative golden file and rejects absolute paths, traversal, and symlinks outside the project. Strict mode returns that relative path, replaces evaluation failures with a generic message, and masks configured model, storage, source, and access-log paths in diagnostic responses.

Embedding applications can use createMcpServer(cwd) to construct a server around a caller-owned transport, or connectMcpServer(transport, cwd) to connect it and receive a closeable handle. serveMcp(cwd) remains the simplest local stdio entry point. A server lazily reuses one RagmirClient per effective configuration for its pinned project root, refreshes it after configuration changes, and closes the active client during shutdown. Request cancellation reaches retrieval, audit, evaluation, security, usage, and bounded resource operations. Native filesystem and LanceDB calls that cannot receive an AbortSignal directly are checked immediately before and after the call. Ragmir does not open an HTTP port, so a network-facing host owns its authentication and authorization boundary.

Teams and synchronized sources

Ragmir has no hosted database or built-in cloud sync by design. Synchronize one source-of-truth folder with an existing tool, such as the Google Drive desktop app or a team script, then let each developer ingest locally. Keep the Ragmir version, configuration, embedding provider, and model aligned so the resulting local indexes remain equivalent.

Sync before rgr ingest, then run rgr audit. A missing, partially synchronized, or extra file in a selected raw or source folder makes that developer's evidence diverge. Synchronize source files, not an actively written .ragmir/storage/ directory.

Monorepos

rgr bases --json
rgr --project-root apps/web search "checkout contract"

Ragmir selects the nearest .ragmir/config.json. Root and nested bases keep separate storage. Generated MCP helpers pin RAGMIR_PROJECT_ROOT, and nested bases use deterministic server names so an agent does not silently query the wrong index.

If a client can see more than one Ragmir server, call ragmir_status and verify knowledgeBaseId before retrieval.

Canonical source: docs/agent-integration.md.

Clone this wiki locally