AIエージェントを絶対に制御する

Free & private — star it on GitHubSync is the project-management and quality-control layer your AI agents work against. Define your specs, rules, decisions, and roadmap once — Sync holds every agent accountable to them, so they build against your architecture instead of hallucinating or repeating past mistakes. When the code moves, stale context gets flagged and the agent has to stop and ask.From install to results in one step. Runs entirely on your machine — nothing is ever uploaded. Currently in alpha · v0.4.5. Building team & cloud features — join the waitlist.MCP server · Desktop app · Built-in terminal · Claude Code · Cursor · Zed · Codex · OpenCodesync — workspaceBad context is how you and your agent drift apartAn agent is only as aligned as what it reads. Point it at stale, scattered context on a large task and it heads off course — confidently, while you're looking the other way.Pattern-matches stale textIt trusts whatever it finds, even when the code moved on months ago.Hallucinates the gapsWhere docs are missing, it invents plausible-but-wrong context.Re-introduces dead decisionsIt re-litigates choices you already overturned, again and again.Get started in secondsThree steps to agent controlNo mandatory terminal commands, no git clone, no long configuration files. Everything happens in the desktop app or with a single click.01Install the desktop appDownload Sync for macOS, Windows, or Linux. One binary, no dependencies — just open it.02Open your project folderPoint Sync at any local repository. It sets up the .sync/ knowledge base automatically — no manual config.03Connect your favorite agentThe Welcome wizard wires up Claude Code, Cursor, Zed, Codex, or OpenCode instantly — and you can run them right inside Sync's built-in terminal. No config files.Latest release v0.4.5 · macOS, Windows & Linux · free and open sourceWorkspaceApi2ASW+DecisionvalidUse queue-based ingestion for uploadsDecouples the API from processing so a slow worker never blocks a request. Supersedes the inline approach.src/server/ingest.tslinked to 2 constraintsConstraintstaleUploads must stay under 10 MBThe file it describes changed in the last commit — Sync flags it for review before your agent trusts it.src/server/upload.tstouched 2h agosync · alpha · a1b2c3dM1 read-only · v1How Sync is differentBuilt to govern agents, not just feed them textAI coding agents are powerful, but without a plan to work against they cause architectural drift, forgotten decisions, and regressions. Memory files and wikis can hold context — they aren't built to keep an agent accountable to your project. Sync is: a mission-control layer above your editor and agent.SyncProject management & quality controlMemory systemse.g. Cursor rulesDocumentationConfluence · NotionPrimary goalControl agent results & enforce qualityProvide simple text context to the agentHuman-readable knowledge sharingStructureStrict, agent-readable constraints & widgetsFree-form text filesUnstructured text, tables, imagesStaleness trackingAutomatic — tied tightly to code changesNone / manualManual audits requiredAgent behaviorMust respect the rules; asks for approval when context goes staleReads notes, but often ignores themUsually can't access it seamlesslyUser experienceVisual Project Hub, widgets, direct oversightSimple text editingExternal web interfaceSync works as a highly effective memory system too — but its real power is governance and quality control.Why SyncOne source of truth you and your agent shareStructured, versioned, and always current — so you and your agent (and the whole team) build the same thing, instead of you cleaning up after it drifts.Project Hub & WidgetsYour mission control for AI agentsNot just notes — a task tracker agents actually work against, with a visual dashboard that shows exactly what's happening in your project right now.A roadmap your agent readsSet milestones, define specs, and the agent pulls what's next from your plan — not its own guess. Blocked work drops out of the queue. You see progress in real time.Widgets you or your agent buildAdd widgets to visualize project health, open questions, stale entries, or any metric. Your agent can create dashboards too — turning Sync into a living operations console.And moreBuilt to keep your shared context honestA validation state you can trustEvery entry is valid, unverified, stale, or invalid — so neither you nor your agent ever builds on context that quietly went wrong.Speaks your languageChoose the language your agent writes docs in. Ready-made presets, plus any BCP-47 tag — so specs, decisions, and docs read naturally for your team.Backlinks & knowledge graphTyped links — depends_on, supersedes, references — built from frontmatter, manual links, and body mentions.Questions & inline commentsDiscussion pinned to exact text. A human answer becomes the canonical, attributed part of the spec.Multi-project workspacesSwitch between projects from the side rail, or group several into a curated workspace and steer them as one.Tied tightly to your codeEvery entry is bound to the files it describes. When the code changes, Sync re-checks the affected knowledge and flags what drifted — no manual audits.Built-in quality gatesWe don't pretend AI never gets it wrongProjects change. Files move. APIs get renamed. Left unchecked, your agent keeps trusting context that quietly went stale — and it won't tell you. Sync's staleness engine ensures it has to.Flags stale context automaticallyWhen a commit touches files a spec or decision describes, Sync marks it stale — pointing at the exact commit and drifted files. No manual audit.Forces the agent to stop and askInstead of silently building on old data, the agent has to surface what went stale and ask you to re-validate it — shutting down the #1 hallucination path.You review and approveNothing is auto-committed. You review the agent's proposed decisions and answer its questions in the Project Hub — you're always in the driver's seat.Who is it forFor anyone who ships code with AIWhether you're building solo or leading a team, Sync makes your relationship with agents predictable — and keeps you in charge.Founders & indie hackersDelegate routine to agents, keep the architect role. Set the direction once — the agent builds against your plan, not its guess.DevelopersStop repeating instructions to your agent. Write a spec once, link it to the code, and every session gets the canonical version — reasoning included.Product managersWrite specs the AI actually reads and executes — not prose that dies in Notion. Link acceptance criteria to milestones. Watch progress happen.And not just software. Because your agent authors custom widgets, the Project Hub adapts to whatever you're building:Trading & automation botsLanguage learningResearch & knowledge basesContent & writingPersonal projectsBe first when the team features landSync is free and open source today. We're building team and cloud features next — drop your email and we'll tell you first. No noise in between.FAQFrequently asked questionsWhat is Sync?+Sync is the project-management and quality-control layer your AI agents work against. You define the project — its specs, rules, decisions, and roadmap — and Sync holds every agent accountable to it, so they follow your architecture and constraints instead of hallucinating or repeating past mistakes. Everything lives in your repository; there is no cloud, no account, and no telemetry.How is it different from a wiki or Markdown docs?+Wikis and Markdown rot silently: nothing tells you when a page no longer matches the code, and your agent usually can't access them seamlessly anyway. Sync stores structured, agent-readable specs and constraints, ties each one to the files it describes, and flags it stale when those files change — so the agent never pattern-matches against out-of-date information.Isn't this just a memory system for agents?+You can use Sync as a highly effective memory system — but that's not the point. Memory tools pass raw text to the agent and hope it's still true. Sync's job is governance: it ties every spec, decision, and constraint to the code it describes, makes the agent respect them, flags them stale when the code changes, and forces it to stop and ask for your approval before acting on old data. The power is control and quality, not just recall.How does it keep me in control of the agent?+You set the plan the agent works against — goals, milestones, and specs ranked by priority. As it works, the agent reads your constraints, updates the roadmap, and files open questions when it hits ambiguity; your answers become canonical. You overturn decisions and set constraints it must respect. Nothing is auto-committed — you review and approve every change. You steer the direction; the agent follows.Where is my data stored?+In your repository. Sync writes plain files under `.sync/` — there is no SaaS backend, no account, and nothing is uploaded. Your knowledge is versioned, reviewable in pull requests, and travels with the code.Which agents and editors work with Sync?+Any MCP-capable client. The Welcome wizard wires up Claude Code, Cursor, Zed, OpenCode, and Codex for you instantly — and you can run them right inside Sync's built-in terminal. For anything else, point your client at the binary with `claude mcp add git-sync -- git-sync mcp`. Sync ships as a desktop app and a CLI, and the binary itself is the MCP server.Is it ready to use?+Sync is currently in alpha. It's usable today and improving fast, but the internal data model can still change between releases. Hit a rough edge? Open an issue on GitHub — issues and pull requests are warmly welcome.Is it free? What's the catch?+Sync is free and open source under the Functional Source License (FSL-1.1-MIT), which reverts to MIT two years after each release — free for all non-competing use. It runs entirely on your machine. Paid team and cloud features are coming later — join the waitlist to be first in line.Stop hoping. Start knowing.Sync is free and open source. Nothing is uploaded — everything lives in your repo, on your machine. Your agent works from reviewed context that matches your architecture, and you approve every change.

※ 著作権に配慮し、引用は冒頭3段落までです。続きは元記事をご覧ください。