factlog
A shared, append-only fact log that enables AI agents to coordinate on a workspace via MCP, CLI, REST, and web interfaces.
README
factlog
A shared, append-only fact log that lets AI agents (and their humans) coordinate on one workspace — a repository, or any bounded domain of agent activity: decisions, invariants, claims on file or uri scopes, questions, findings, and handoffs — persisted in SQLite, served over MCP (stdio and streamable HTTP), a CLI, a REST ingest endpoint, and a live web view.
- Protocol semantics: factlog-rfc-v0.1.md
- Implementation spec: factlog-build-spec.md
- Context-discipline addendum (hooks, runner): factlog-context-discipline.md
Install
Node ≥ 20.
npm install && npm run build && npm link # puts `factlog` and `factlog-mcp` on PATH
npm test # full suite, incl. the 10-process concurrency gate
Initialize a workspace
cd your-repo
factlog init # creates .factlog/factlog.db, adds it to .gitignore
Connect an agent (MCP)
Add to the repo's .mcp.json:
{ "mcpServers": { "factlog": { "command": "npx", "args": ["factlog-mcp"] } } }
The server exposes factlog_brief, factlog_post, factlog_query,
factlog_changes, factlog_claim, factlog_supersede, factlog_release,
and factlog_resolve. Set FACTLOG_SESSION / FACTLOG_AGENT so writes are
attributed to the right session.
The same 8 tools are served over MCP streamable-HTTP at
http://127.0.0.1:4711/mcp by factlog serve (and on the workspace unix
socket .factlog/factlog.sock, mountable into containers). Identity comes
from x-factlog-agent/-session headers — or, hardened, from an actor token
(factlog token mint) presented as Authorization: Bearer, which stamps the
actor server-side, marks writes x-actor-verified, and can carry write-scope
ACLs plus an external taint (decision/invariant posts then land pending
until approved in the web UI). Set require_stamped_actors in
.factlog/policy.json to reject unstamped writes.
CLI
factlog brief # deterministic, budgeted context pack
factlog post decision --body "..." --scope 'src/auth/**'
factlog post note --body "..." --scope 'topic://finances/**' # uri scopes (RFC §3.4)
factlog claim --scope 'job://daily-digest' --task "..." # doubles as a run-lock; exit 2 on conflict
factlog release <id> --commit <sha>
factlog watch # tail -f the log
factlog serve # web view :4820 + daemon :4711 (/mcp, /ingest, hooks) + unix socket
factlog token mint --agent whatsapp --origin external --write-scope 'channel://**'
factlog bridge # fenced brief block in CLAUDE.md
factlog export # full log as NDJSON (no lock-in)
factlog help lists the rest, including the hook-facing commands
(check-claim, session-status, next-task).
Operational events (RFC Appendix C)
CI, incident, and deploy tooling POSTs distilled judgments to
http://127.0.0.1:4711/ingest: incidents and red test suites become findings
auto-resolved by their closing transition; deploy freezes become blocking
invariants with meta["x-gate"]="block", which the PreToolUse hook enforces
mechanically — agents cannot edit a frozen scope — and meta["x-expires"]
expires them on schedule.
Hooks (enforced discipline for Claude Code)
demo/hooks/ ships a ready-made settings.json plus scripts: brief injected
on session start, edits blocked unless the session holds a covering claim,
stop gated on releasing claims and posting a handoff. Install per
demo/hooks/README.md.
Demo
The two-agents-one-repo walkthrough — parallel claims, a conflict flashing
red on the live timeline, a blocking question resolved across sessions, a
superseded decision — is scripted in demo/demo.md. Seed a
non-trivial log first with npm run seed.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。