Outreach API MCP
joethefisher/outreach-api-mcp
MCP server read-only para Outreach — prospects, contas, sequências e mais
Ver no GitHubServidor MCP independente para integração com a plataforma Outreach, expondo 21 ferramentas de leitura para consultar prospects, contas, sequências, templates, snippets, tarefas, mails e logs de auditoria. Conecta Claude Desktop ou qualquer cliente MCP direto à sua workspace Outreach via OAuth, sem afiliação oficial.
Instalação
npx convez-ai joethefisher/outreach-api-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Consultar e analisar dados de prospects e contas em tempo real durante workflows de vendas
- Auditar sequências de email e templates sem sair do Claude
- Revisar logs de auditoria e atividades da workspace para conformidade
Veredito

Integração sólida e bem-revisada (v0.1.2) que expõe a API Outreach de forma segura como MCP. Read-only por design e com autenticação OAuth robusta — ideal para quem já usa Outreach e quer consultá-lo via Claude sem ferramentas de escrita.
README
A tradução para o português deste README ainda está na fila.
Outreach API MCP
A Model Context Protocol server for Outreach. You give it your own Outreach OAuth credentials, an MCP client like Claude Desktop talks to it over stdio, and you get 21 read-only tools for querying prospects, accounts, sequences, templates, snippets, tasks, mailings, and audit logs from your own workspace.
Independent integration. Not affiliated with, endorsed by, or sponsored by Outreach. You comply with Outreach's API terms; this code just speaks the protocol.
Status
v0.1.2. The MCP server is in outreach-worker/ and runs over stdio. Read-only by design: no POST, PATCH, or DELETE, and no write scopes get requested. CI runs on Node 20 and Node 22 against 290 unit and integration tests. Three review passes have landed: the initial scaffold review, the v0.1.1 blocker-fix verification, and the v0.1.2 fast-follows (correctness, security hardening, test coverage, and design cleanup — see commit history on main for the per-finding trail).
Prerequisites
Two things:
- Node.js, version 20 or newer.
- An Outreach OAuth application in your own Outreach workspace. You register this once and reuse it forever.
Registering an Outreach OAuth app
You do this once per workspace. The server uses these credentials to mint per-user refresh tokens, and credentials never get shared across users of your workspace.
The steps:
- Sign in to Outreach as a user with admin permission for API integrations.
- Open Settings (gear icon), then Integrations, then API Access, then "Your applications."
- Click the button to create a new app.
- Fill in the form:
- Name: anything descriptive, like
outreach-api-mcp (local). - Description: optional.
- Redirect URIs: add
http://127.0.0.1:8765/callback. If you later change the port viaOUTREACH_OAUTH_REDIRECT_PORT, the URI here has to match exactly. - Scopes: pick every scope listed in
outreach-worker/src/auth/scopes.ts. All of them end in.read. If you grant fewer, the tools that need a missing one will return ascopeMissingenvelope and the agent will tell you which.
- Name: anything descriptive, like
- Save the app.
- Copy the Application ID into
OUTREACH_CLIENT_IDand the Application Secret intoOUTREACH_CLIENT_SECRET. Outreach will not show the secret again, so put it somewhere safe.
A couple of gotchas. If the API Access page is hidden in your UI, an admin needs to grant you "Manage API Integrations" first. And the redirect URI has to match exactly, down to scheme, host, and port. http://127.0.0.1 (not localhost) is the right value; Outreach accepts it as loopback per RFC 8252.
Install
Two paths. Pick whichever feels more natural.
From a tagged release tarball (no clone required; treats this like a normal npm dependency):
npm install -g https://github.com/joethefisher/outreach-api-mcp/releases/download/v0.1.2/outreach-api-mcp-0.1.2.tgz
That gives you the outreach-api-mcp binary on your PATH plus the bundled dist/. Skip ahead to the bootstrap section.
From source (if you want to read the code, contribute, or run on an unreleased commit):
git clone https://github.com/joethefisher/outreach-api-mcp.git
cd outreach-api-mcp/outreach-worker
npm ci
npm run build
Bootstrap
# Put your app credentials in ../.env first (the bootstrap script reads it).
cp ../.env.example ../.env # then edit OUTREACH_CLIENT_ID / SECRET
# One-time browser consent flow. Captures the refresh token.
npm run bootstrap:oauth
What bootstrap:oauth does, in order: opens your browser, walks you through the Outreach consent screen, captures the redirect on a local loopback HTTP server, exchanges the code for a refresh token, persists it to the on-disk cache, and prints an .env block on success. Copy that block into your .env (or wherever your MCP client sources env from). That is the whole setup.
Wiring into an MCP client
Once you have all three credentials, point your MCP client at the built server. The bootstrap step is only needed once; from this point on the credentials live in your MCP client's config and the on-disk token cache.
For Claude Desktop, find claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; Windows: %APPDATA%\Claude\claude_desktop_config.json) and add an entry:
{
"mcpServers": {
"outreach": {
"command": "node",
"args": ["/absolute/path/to/outreach-api-mcp/outreach-worker/dist/index.js"],
"env": {
"OUTREACH_CLIENT_ID": "...",
"OUTREACH_CLIENT_SECRET": "...",
"OUTREACH_REFRESH_TOKEN": "..."
}
}
}
}
Fully quit and reopen Claude Desktop (a window reload is not enough — the MCP transport only restarts on app launch). The 21 tools below will show up in the tool picker. The server reads its configuration from the env you provide in this config block, not from any .env file on disk.
For other MCP clients, the contract is the same: spawn node dist/index.js as a subprocess with the three env vars set, and talk to it over stdio.
Configuration
Everything is environment variables. .env.example has the full list. The minimum:
| Variable | Purpose |
|---|---|
OUTREACH_CLIENT_ID | OAuth client ID from your app |
OUTREACH_CLIENT_SECRET | OAuth client secret from your app |
OUTREACH_REFRESH_TOKEN | Long-lived refresh token from bootstrap:oauth. Seeds the on-disk cache on first run; after that the cache is the source of truth. |
Available MCP tools
21 tools across five categories: 19 capability tools and 2 escape hatches. Schemas are in outreach-worker/src/index.ts if you want to read them directly.
| Category | Tools |
|---|---|
| Prospects and Accounts | searchProspects, getProspectProfile, searchAccounts, getAccountProfile |
| Sequences | searchSequences, getSequenceProfile, getProspectSequenceHistory, analyzeSequencePerformance, compareSequences |
| Templates and Snippets | searchTemplates, getTemplate, searchSnippets, getSnippet |
| Activity and Audit | getOpenTasks, getRecentMailings, getTeamRoster, getUserActivity, getAuditLog |
| Drafting | draftEmail (returns a context bundle; never sends mail) |
| Escape hatches | outreachQuery, outreachGetById |
Read-only invariant
The part I want to be very explicit about. The server is read-only. No POST, no PATCH, no DELETE. The HTTP client hardcodes GET on every call, and there is no path through any tool that mutates Outreach state. Only .read OAuth scopes are requested. Even draftEmail only returns context; the model composes the email in chat for you to paste into Outreach yourself.
If you find a tool that violates this, that is a bug. Open an issue.
Security model
SECURITY.md has the full threat model. Briefly:
- No telemetry. No outbound network calls beyond Outreach API endpoints.
- All logging goes to stderr only. Stdout is reserved for MCP JSON-RPC frames, and mixing them up would corrupt the transport, which is why it is enforced by lint and asserted in tests.
- OAuth tokens live in an on-disk cache with
0600permissions, in$XDG_CONFIG_HOME/outreach-api-mcp/. The parent directory is0700. - Refresh tokens rotate; the latest one always gets written back to cache.
- Tokens, PII, and message bodies are never logged. The structured logger redacts on the way out.
- Scope minimization: only the 19 read scopes that the tools actually need.
Standards
This codebase commits to a quality bar in STANDARDS.md. PRs get reviewed against it. The short version is strict TypeScript everywhere, integration tests for tool blocks, and no features the task did not request.