slivingdoc

slivingdoc is an open-source MCP server and command-line tool that gives AI agents and people one shared directory of durable notes. The notebook is stored in an S3-compatible bucket, but agents work with ordinary UTF-8 files in a local directory. That makes it useful for context that is too loose for a database schema, such as investigation progress, project decisions, customer notes, runtime instructions, or reports from a fleet of workers. Sessions in Claude Code, Codex, or any other MCP host can pull the same notebook, add what they learned, and publish changes without introducing a separate editor or proprietary document format. People can join the same workflow from a terminal and any text editor, even without an MCP host.

The interface is deliberately small. The MCP server exposes notes_pull and notes_commit, and the CLI offers matching pull and commit operations. Pull reads the accepted notebook state and writes the whole notebook into a checkout directory. Agents or humans then edit those files with the tools they already use. Commit compares the current directory with its remembered baseline, incorporates the latest accepted state, and publishes the change. A checkout must be pulled before its first commit, which prevents a process from publishing a guess about state it has never seen. Successful results report the accepted generation and per-file insertion and deletion counts. Domain errors return a retry verdict and a suggested next step. There is no background daemon or hidden synchronization between operations.

Concurrency is handled with Git semantics behind the scenes and an optimistic publication model in S3. The accepted state is named by a small current manifest that points to immutable packs. Uploading a pack alone doesn't publish it, since the conditional manifest swap is what makes a proposal accepted. When two writers race, the bucket's conditional replacement accepts one update for the ETag they both observed. The other writer downloads the new state, merges what it missed, and retries with a fresh proposal. Non-overlapping work can therefore land without a lock, lease, or shared clock. If both writers changed the same lines, slivingdoc returns a CONTENT_CONFLICT result and writes standard conflict markers into the affected files rather than silently choosing a winner.

Conflicts remain ordinary text that an agent or person can resolve with the same editor used for the note. Structured errors identify every affected file and the relevant marker ranges, then the caller can remove the markers and commit again. Complete conflict blocks aren't accepted into the shared notebook. The storage layer also creates checkpoints after a configurable number of incremental packs, so a new reader can fetch a complete-state checkpoint and a short tail rather than replaying an ever-growing history. Checkpoint compaction happens asynchronously, doesn't block writers, and leaves accepted state untouched if a checkpoint attempt fails.

The design supports several practical patterns. Coding agents in different harnesses can share project context from one session to the next. Short-lived CI jobs, scheduled workers, and disposable sandboxes can recover memory after their local disks disappear. Production agents can keep customer needs, investigation state, and prior attempts in plain text. A large fleet can write reports into separate directories and let a human pull the combined notebook. Teams can also expose selected paths as read-only to agent processes, which works for instructions that a person updates through a fully writable process. A fixed workspace root lets people and agents share a visible directory, while the default MCP server uses a private scratch directory and treats the bucket as the durable boundary.

Setup requires a supported Node.js runtime, an existing bucket, and credentials available through the normal AWS chain. The project documents AWS S3, Tigris, MinIO, and other compatible stores that support conditional writes. Its npm launcher downloads the native binary for Linux, macOS, or Windows, verifies its checksum, and runs it, while native releases and an offline SeaweedFS example are also available. The included Terraform module can provision a bucket and a least-privilege IAM user for one notebook. The main limitation is intentional. The notebook accepts clean UTF-8 text, not binary files, and synchronization happens only when pull or commit is called. slivingdoc is free under the MIT license, though users still provide and pay for their own storage. It is a young project with a small community, but its public architecture, documentation, Terraform module, and CLI make the operating model inspectable.