Docs · integrations

Integrations

CLI + Skill, MCP, live streaming

Surfaces

Causalist ships three ways to invoke it, all backed by the same core graph:

Web app

Paste a GitHub URL at causalist.xyz and Claude maps the repo. Hand-curated demos load instantly; live analyze of any public repo runs when you add your Anthropic key in Settings.

CLI + Skill (preferred for Claude Code)

npm i -g causalist-cli
causalist install        # drops SKILL.md into ~/.claude/skills/causalist/
causalist pair <code>    # one-time browser pair

The CLI ships eleven graph-query subcommands (causalist node, blast, tests, writers, path, …) — JSON-by-default when piped, plain text in TTY, exit codes 0 ok / 1 ok=false / 2 fatal-stderr. causalist install drops a Claude Code Skill at ~/.claude/skills/causalist/SKILL.md so the agent auto-discovers when to use which command.

This is the path Anthropic recommends in Code Execution with MCP: code over tool-call JSON. Tool descriptions don't pollute context; only the chunks the agent actually needs come back.

MCP server (for non-CLI clients)

claude mcp add causalist -- npx -y causalist-mcp@latest --session ABC123

For Cursor, Claude.ai web, or any client that doesn't have a shell. Same eleven tools exposed over stdio. Pair codes come from /pair.

The eleven tools (CLI + MCP)

CLIMCPWhat it does
causalist node query_nodeNode metadata, layer, summary
causalist neighbors get_neighborsIn/out edges + verified flag
causalist path find_pathShortest causal path
causalist layer find_nodes_by_layerFilter by semantic layer
causalist blast blast_radiusWhat transitively depends on a node
causalist tests affected_tests"3 tests not 300" — reachable tests
causalist writers find_writersEvery node that writes to a target
causalist similar similar_nodesStructurally similar nodes
causalist topo topo_orderTopological layering of a subgraph
causalist verify verify_edgeConfirm an edge exists
—create_project(MCP-only) push a project to the paired browser
causalist info—Capabilities manifest with active session

Live streaming

When Claude Code is paired with the browser, its PostToolUse hook posts tool-use events to /api/ingest/. The browser viewer subscribes via Server-Sent Events at /api/stream/ and highlights nodes in real time — open the repo's graph in another tab and watch Claude work.

Pairing is how the browser and local Claude session agree on a sessionId without requiring a user account.

Plan-mode agent runs (web)

The Agent tab in the right-side panel exposes two endpoints:

  • POST /api/agent/run — SSE. Body { plan, repo, branch, selectedNodeIds, nodePathMap, apiKey }. The user types a plan in plain English ("audit these files for security issues and fix any injection vulnerabilities"); the server streams file_loaded, finding, patch, done, error events as Claude Opus 4.7 carries it out against the selected files.
  • POST /api/agent/push-pr — uses your GitHub OAuth cookie to create a branch via the Git Tree+Commit API and open a real pull request from the patches.

Analyze API

POST /api/analyze is a public SSE endpoint. Pass Authorization: Bearer and a JSON body {owner, repo, commit, tree, files?}; receive agent events as each of the four pipeline agents (Structure, Dependency, Semantic, Oracle) completes, then a done event with the final graph.