Experimental software. APIs, schemas, and behaviors change without notice. Use at your own risk. Contributions welcome — expect rough edges.
Autonomous work orchestrator for AI agents across multiple code repos. Local-first, open source. Hosted version coming at rosary.bot.
Rosary structures work as beads — small, trackable units stored in each repo via Dolt. A reconciliation loop scans for ready beads, dispatches AI agents (Claude, Gemini) to execute them in isolated workspaces, verifies the results, and syncs status to Linear for human review.
The human reviews 5-10 feature PRs a day. The agents handle the atoms.
graph LR
A[Scan repos] --> B[Triage & prioritize]
B --> C[Dispatch to agent]
C --> D[Agent works in<br/>isolated workspace]
D --> E{Verify}
E -->|pass| F[Close bead]
E -->|fail| G[Retry with backoff]
E -->|deadletter| H[Flag for human]
G --> C
stateDiagram-v2
[*] --> open
open --> dispatched: agent assigned
dispatched --> verifying: agent completes
verifying --> done: checks pass
verifying --> open: retry
verifying --> blocked: deadlettered
blocked --> open: human /resume
Work items live in each repo as beads — an AI-native issue tracker backed by Dolt (version-controlled SQL). Rosary reads and writes beads directly over MySQL, no CLI shelling.
Beads are organized into threads (ordered progressions of related work) and decades (ADR-level groupings) via the BDR harmony lattice.
# Beads are managed via rosary's MCP tools or CLI:
rsry bead create "Fix auth bug" --priority 1 --type bug --files src/auth.rs
rsry bead list
rsry bead search "auth"
rsry bead close rsry-abc123task build # requires Task (taskfile.dev)
task test
# Register repos to watch
rsry enable ~/code/my-app
rsry enable ~/code/my-lib
# See what's ready
rsry status
# Dry run — see what would be dispatched
rsry run --once --dry-run
# Real run — dispatch agents, verify, close
rsry run --once --concurrency 3
# Continuous loop
rsry runUse
task build/task testinstead of rawcargo— the Taskfile setsPKG_CONFIG_PATHfor the fuse-t dependency via ley-line.
Rosary exposes 24 tools as MCP. Any AI agent or human with an MCP client can scan beads, dispatch work, manage threads, and track progress.
# Add to Claude Code (one-time)
claude mcp add -s user rsry -- rsry serve --transport stdio
# Or run as HTTP server
rsry serve --transport http --port 838324 tools across five categories:
| Category | Tools |
|---|---|
| Beads | rsry_bead_create, rsry_bead_update, rsry_bead_search, rsry_bead_comment, rsry_bead_close, rsry_bead_link |
| Status | rsry_status, rsry_list_beads, rsry_scan, rsry_active |
| Dispatch | rsry_dispatch, rsry_run_once, rsry_decompose, rsry_pipeline_upsert, rsry_pipeline_query, rsry_dispatch_record, rsry_dispatch_history |
| Workspaces | rsry_workspace_create, rsry_workspace_checkpoint, rsry_workspace_cleanup, rsry_workspace_merge |
| Hierarchy | rsry_decade_list, rsry_thread_list, rsry_thread_assign |
See docs/CONFIGURATION.md for the full reference.
# ~/.rsry/config.toml
[[repo]]
name = "my-app"
path = "~/code/my-app"
[[repo]]
name = "my-lib"
path = "~/code/my-lib"
[linear]
team = "ENG"
[backend]
provider = "dolt"
path = "~/.rsry/dolt/rosary"
[compute]
backend = "local" # or "sprites" for remote containersAgents run in isolated workspaces (jj preferred, git worktree fallback). The compute backend is pluggable:
| Provider | What | Config |
|---|---|---|
local |
Host subprocess (default) | none |
sprites |
sprites.dev containers | SPRITES_TOKEN env var |
Bidirectional sync — beads are source of truth, Linear is the UI. Threaded beads sync as sub-issues.
rsry sync --dry-run # preview
rsry sync # push + pull + reconcileWebhooks for real-time updates: rsry serve --transport http exposes /webhook.
After an agent completes, rosary runs tiered checks:
- Did it commit something?
- Does it compile?
- Do tests pass?
- Does the linter approve?
- Is the diff a reasonable size?
Failed checks trigger retry with backoff. After 5 failures or 3 regressions, the bead is deadlettered for human attention.
Rosary manages its own development. It scans its own repo, dispatches agents to fix its own bugs, and verifies the results. The plumbing works — proving it at scale is ongoing.
See docs/ARCHITECTURE.md for the full technical picture, docs/CONFIGURATION.md for config reference, and docs/glossary.md for terminology (beads, decades, threads, triage, etc.).
task build # debug build with fuse-t support
task test # run tests
task lint # fmt + clippy
task all # fmt + check + lint + testPre-commit hooks enforce cargo fmt and cargo clippy on every commit.