Provider-agnostic by contract
A board is a contract; Jira, Trello, and a built-in local board are implementations. Version control is a contract; GitHub, Bitbucket, and GitLab (incl. self-hosted) are adapters. Adding one changes nothing else.
Ryvem is a provider-agnostic MCP server with a temporal memory graph and a deterministic agent harness — shipped as one self-contained binary. It conducts your agent through refine, branch, implement, open PR, and move the card, grounded in your repository's own conventions.
Install — one line, any OS
curl -fsSL https://ryvem-mcp.viniciusborges.dev/install.sh | bash
irm https://ryvem-mcp.viniciusborges.dev/install.ps1 | iex
Ryvem supplies the board, the branch memory, the repository knowledge, the pull-request plumbing, and the guardrails. Your agent supplies the intelligence; Ryvem conducts it.
A board is a contract; Jira, Trello, and a built-in local board are implementations. Version control is a contract; GitHub, Bitbucket, and GitLab (incl. self-hosted) are adapters. Adding one changes nothing else.
One SQLite knowledge graph with full-text search — typed nodes and edges on a structural axis, and a temporal axis where nothing is deleted. Every update supersedes; asOf and history make the past first-class.
Ryvem records the exact commit it read at. On the next pull it diffs against that anchor server-side: targeted re-reads, new reads, and every business rule whose evidence file vanished flagged for verification. You never re-read the whole repo again.
Expert roles — PM, Tech Lead, Architect, Engineer, DevOps, Security — move work through loop state machines persisted in the graph. Gates demand evidence: Ryvem runs your test, lint, and typecheck and passes only on a real exit 0.
The strong model orchestrates; cheaper tiers execute its plan. Any provider your client can address is just a provider/model string. Ryvem expresses this as configuration and delegation briefs — the client always owns model selection.
A budgeted repo map, a context ledger that replaces held artifacts with a pointer, output shaping that caps every flood, a blob-hash cache, and cache-stable briefs. The numbers below come from a zero-inference benchmark on real inputs.
Emit the entire knowledge graph as one self-contained HTML file — a force-directed view coloured by node type, a markdown reader, instant search, and a history toggle that walks the supersede chain as a timeline. No server, no tokens, no external requests.
Windows, Linux, and macOS (Apple Silicon and Intel) — the memory graph is embedded, so there is no database to run, no service to host, no dependency to install at the destination. One file next to a .env, or Node ≥ 22.13 if you prefer the JavaScript.
These measure Ryvem's own engineering — the context it hands the model is dramatically smaller and better-targeted. They do not measure a model solving a task.
~1.4 MB of TypeScript rendered as a 7.7 KB ranked symbol skeleton fitted to an exact token budget.
A 200-file repository where a pull changed four files re-touched only three — and flagged the rule whose evidence was deleted.
A 488 KB command flood clipped to 19.6 KB, carrying a "run a narrower query" nudge instead of the dump.
After editing one of 30 files, only the changed git blob is re-parsed; the other 29 are served from the graph cache by blob hash.
Repeated reports and ABAB oscillation are diverted into a re-plan by deterministic pattern, not inference.
Synthetic benchmark, zero model inference — every figure is a deterministic count or ratio (not a wall-clock timing), reproducible with bun bench/synthetic.ts.
Run the one-liner. The installer detects your platform, checksum-verifies the binary, puts ryvem-mcp on your PATH, and registers it with any agent CLI it finds.
The setup wizard configures Jira or Trello with the token URL and scopes shown for each. Skip it and Ryvem boots zero-config on a private local board that works offline.
Point your agent at Ryvem and run start-session. It learns your stack, follows your conventions, and remembers what it learns across sessions and repositories.
Read the getting-started guide: install per OS, configure a board, and register Ryvem with Claude Code, Cursor, Antigravity, or Codex.