AgentLink MCP Server
AgentLink MCP server enables cross-repo contract negotiation for coding agents, exposing a stdio API to list agents, start conversations, send/read structured messages, update/accept contracts, and close coordination sessions across Claude Code, Codex, Copilot, and other MCP-capable harnesses.
README
AgentLink
AgentLink is a local-first coordination layer for multiple coding agents — Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Gemini CLI, and similar harnesses — working across related repositories.
The product goal is not generic agent chat. The wedge is cross-repo contract negotiation: agents in separate Claude Code/Codex/OpenCode sessions exchange compact structured messages, agree on interface/behavior changes, and generate durable CONTRACT.md handoff files each repo can implement against.
Install
Run without installing:
npx agentlink doctor
npx agentlink setup --harness all
Install globally:
npm install -g agentlink
agentlink doctor
agentlink setup --harness claude-code
For MCP-capable harnesses, the package also exposes agentlink-mcp as a stdio server.
Demo

The README animation above shows two local workspaces negotiating an API contract, recording approvals, accepting the contract, and syncing .agentlink/CONTRACT.md to the peer repo.
The source recording is also checked in as an asciinema v2 cast:
asciinema play demos/agentlink-demo.cast
Regenerate both demo assets locally with:
python3 scripts/record-demo.py
python3 scripts/cast-to-gif.py
MVP
- Discover active coding-agent sessions from tmux panes across common harnesses: Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Gemini CLI, Kimi, and unknown future MCP-capable CLIs.
- Let a user select a target repo/session.
- Start a structured coordination conversation.
- Persist all messages to
.agentlink/conversations/*.jsonl. - Generate/update
.agentlink/CONTRACT.md.
Commands
npm install
npm run agentlink -- init
npm run agentlink -- list
npm run agentlink -- start --topic "OAuth contract change" --target api-agent
npm run agentlink -- start --topic "Add account summary endpoint" --target api-agent --template api-change
npm run agentlink -- start --topic "Bounded API negotiation" --max-rounds 6 --required-approvals 2
npm run agentlink -- send --from web-agent --role assistant --body "Please add expiresAt."
npm run agentlink -- send --body "Please review the stored message." --deliver-to %2
npm run agentlink -- read
npm run agentlink -- replay
npm run agentlink -- replay --format json
npm run agentlink -- context
npm run agentlink -- context --format json
npm run agentlink -- doctor
node dist/cli.js doctor --format json
npm run agentlink -- setup
npm run agentlink -- setup --harness claude-code
npm run agentlink -- setup --harness copilot
npm run agentlink -- setup --harness gemini
npm run agentlink -- setup --harness stdio --format json
npm run agentlink -- version
npm run agentlink -- ship-check
npm run agentlink -- ship-check --format json
npm run agentlink -- launch-brief
npm run agentlink -- launch-brief --format json
npm run agentlink -- demo --peer ../peer-repo
npm run agentlink -- demo --peer ../peer-repo --format json
npm run agentlink -- approve --from api-agent
npm run agentlink -- contract --status Accepted
npm run agentlink -- contract --status Proposed --set-section "API Surface" --content "- [x] Endpoint: GET /accounts/:id/summary"
npm run agentlink -- contract --status Proposed --sync-to ../peer-repo
npm run agentlink -- status
npm run agentlink -- end
node dist/mcp/server.js
agentlink-mcp is a stdio MCP server exposing the same local bus primitives to MCP-capable coding harnesses: list agents, start conversations, send/read messages, update/accept contracts, and close conversations.
Conversation history is append-only JSONL under .agentlink/conversations/.
read prints recent messages for the latest/open selected conversation; replay
prints the full append-only timeline, including conversation start metadata,
messages, approvals, and close events. Use replay --format json when another
harness needs structured timeline state.
context prints a compact repo fingerprint for handoffs — workspace path,
git branch/commit/dirty-file count, and package scripts — without including source
content or full file lists. Use --format json when another tool needs structured
metadata. doctor checks Node/npm scripts, .agentlink workspace state, current
contract/store health, tmux agent visibility, and the MCP build artifact so local
setup issues are visible before dogfood/demo work; use node dist/cli.js doctor --format json
after building when a harness or CI gate needs machine-readable readiness checks without
npm script banners contaminating JSON. setup prints deterministic
local install, stdio MCP server, harness setup, and agent-prompt instructions;
use --harness <stdio|claude-code|codex|copilot|opencode|gemini|all> and --format json for
machine-readable setup data. ship-check is a read-only launch-readiness gate
for final QA; it verifies package metadata, bins, npm package file allowlist,
README command/positioning coverage, and build artifacts, while explicitly
preserving the boundary that publishing or public launch requires human approval.
launch-brief prints the
final human approval artifact: product thesis, verification commands, demo commands,
launch artifacts, CEO decisions needed, and the no-publish/no-launch-without-approval boundary.
version prints the installed package version so harness configs and smoke tests
can confirm the expected AgentLink build is on PATH.
demo --peer <repo-path> runs a deterministic local two-repo API-contract
negotiation smoke: it creates an append-only conversation, records producer and
consumer messages, records two approvals, marks the contract Accepted, closes the
conversation, and syncs the same .agentlink/CONTRACT.md to the peer workspace.
Use --format json for machine-readable demo checks.
Starting a conversation creates or refreshes .agentlink/CONTRACT.md; init
leaves an existing contract untouched. start --template <template> creates a
focused, deterministic negotiation checklist. Available templates are
api-change, event-contract, db-migration, and frontend-backend; omitting
the option preserves the generic contract. start --max-rounds <count> blocks
further bus messages after a bounded negotiation, and start --required-approvals <count> makes contract --status Accepted fail until enough participants have
run approve --from <participant>. contract --status <status> advances the
local contract state through Draft, Proposed, Accepted, Blocked, Implemented,
or Verified. contract --sync-to <repo-path> copies the current CONTRACT.md
into another repo's .agentlink/CONTRACT.md so both sides have the same durable
agreement artifact. contract --set-section <heading> --content <markdown>
deterministically replaces an existing section or inserts a new one before Status,
so agents can merge concise contract updates without regenerating the whole file.
send only appends by default. With
--deliver-to, AgentLink appends first, then types the structured JSON message
into the selected pane, verifies it is visible, and only then presses Enter.
A tmux delivery failure does not remove the stored message.
Design stance
- Local-first.
- Structured bus is source of truth.
- tmux pane messaging is notification/bridge, not durable storage.
- Repo contexts stay isolated; agents exchange summaries/contracts only.
- Human is founder/approver, not message router.
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。