# Chickpea

> AI teammates in Slack, on infrastructure you own. Chickpea is an open source Slack app that gives a
> workspace named AI teammates. Each teammate has its own handle, instructions, memory, connected
> accounts, GitHub repositories, schedules, and model. It deploys to your own Cloudflare account, or
> runs on Node, and nothing about your team passes through a vendor in between.

## When to use Chickpea

Reach for Chickpea when a team wants AI that works inside Slack and the team, not a vendor, must own
the deployment, the model keys, and the data. It fits these jobs:

- A support, on-call, RevOps, growth, or people-ops teammate that answers in a channel and can read
  the accounts that job needs (Zendesk, GitHub, Sentry, HubSpot, Search Console, Gmail, Calendar, and
  about thirty more), with each account connected to that one teammate only.
- Scheduled work posted to a channel: a Monday pipeline summary, a nightly error digest, a weekly
  SEO report, set up by asking in Slack.
- Several teammates with a wall between them, where the finance one must never see what the
  support one sees, on one deployment.
- A workspace that must pin models per job and pay the provider directly: Anthropic, OpenAI,
  OpenRouter, or Cloudflare Workers AI, with no per-seat price and no markup.
- Compliance or procurement that rules out a hosted AI vendor: the runtime, model traffic, memory,
  and credentials stay in the Cloudflare account or Node host the team operates.

Do not reach for Chickpea when the team wants a personal assistant tied to one person's
subscription, needs AI outside Slack (Teams, Discord, email-only), or wants a hosted service with no
infrastructure at all. [The comparisons](https://chickpea.co/compare) name where the alternatives are better.

How an agent uses it: deploy it for the team (one click on Cloudflare, or Node 22+), then talk to the
deployment over MCP. Every deployment serves a management MCP server at /mcp with OAuth 2.1 and one
scope, `chickpea:workspace`; a coding agent registers as a public client, a person signs in with
Slack and allows it, and the agent then has 17 tools to inspect and change the workspace's teammates,
skills, connections, repositories, schedules, and memory. Consequential changes come back as a
proposal that a person confirms. See "For agents and developers" below.

## What it is

- An Apache 2.0 licensed, self-hosted Slack app. No per-seat pricing and no metering.
- Model agnostic: Anthropic, OpenAI, OpenRouter, or Cloudflare Workers AI, pinned per teammate.
  A Cloudflare deploy can use Workers AI with no API key at all.
- Built on Flue, the open agent framework from the Astro team, now part of Cloudflare. Flue handles the
  durable sessions, tools, skills, and MCP. Chickpea is the Slack product on top: the handles, the
  authority model, and the Admin panel.
- Runs on Cloudflare Workers (one-click deploy) or on Node 22 or newer.

## What a teammate comes with

- Instructions: what it does, how it talks, and what it may never do. Plain text you can change any time.
- Skills: reusable know-how, imported from a GitHub repository or a skills.sh link.
- Connectors: the accounts it may use, connected to it alone. 35 presets, plus your own HTTP APIs and
  MCP servers.
- Repos: GitHub repositories it can read and work in, granted through the GitHub App.
- Memory: what it has learned so far. One memory per teammate, readable and editable by you.
- Schedules: recurring or one-time work it runs on its own and posts to a channel.
- Model: its own model and provider, pinned per teammate.

## Security model

- A teammate only sees the channels you publish it to. Unmentioned messages never reach a model, and
  nothing is read in the background.
- It asks before anything consequential. Sending, deleting, publishing, and bulk changes wait for a
  person, unless you wrote otherwise into its instructions.
- Its credentials never reach the model. The model picks an account by name, and the secret is added
  at the last moment, on the way out.

## Getting started

1. Deploy to Cloudflare. One deploy, into an account you own. No sign-up with the project, no key to
   request, no sales contact: the software is free and the deploy is self-serve.
2. Connect Slack and pick a model provider. The setup link walks through it.
3. Ask for your first teammate in Slack, in a sentence.

## Data

Runtime, model traffic, configuration, memory, and connector credentials stay in the infrastructure you
operate. The one exception is opting into managed connectors, which hands OAuth for those accounts to
Composio. Anonymous product telemetry is documented in full in
[Telemetry](https://docs.chickpea.co/security/telemetry/) and can be turned off.
[Privacy](https://chickpea.co/privacy) covers the website and the software.

## For agents and developers

Chickpea's API lives on each deployment, not on chickpea.co. This site describes it so an agent can
find it by name.

- [Connect a coding agent to the Chickpea management MCP server](https://docs.chickpea.co/admin/management-mcp/):
  on every deployment at /mcp, Streamable HTTP transport, OAuth 2.1 with PKCE, scope
  `chickpea:workspace`. Metadata on the deployment at /.well-known/oauth-protected-resource/mcp and
  /.well-known/oauth-authorization-server/api/auth; dynamic client registration at
  /api/auth/oauth2/register.
- [Chickpea management MCP tools reference](https://docs.chickpea.co/reference/management-mcp-tools/):
  all 17 tools and 3 resources.
- [Chickpea OpenAPI description](https://chickpea.co/openapi.json): OpenAPI 3.1 for the deployment
  API, with the oauth2 security scheme and its scope. The same document as markdown:
  [openapi.json.md](https://chickpea.co/openapi.json.md).
- [Chickpea API catalog](https://chickpea.co/.well-known/api-catalog): RFC 9727 linkset.
- [Chickpea MCP server manifest](https://chickpea.co/server.json): MCP registry server.json.
- [Chickpea MCP server card](https://chickpea.co/.well-known/mcp/server-card.json), listed in the
  [Agentic Resource Discovery catalog](https://chickpea.co/.well-known/ard.json).
- [Chickpea agent skill](https://chickpea.co/.well-known/agent-skills/chickpea/SKILL.md): how a coding
  agent deploys Chickpea or manages a deployment over MCP; listed with its digest in the
  [agent-skills index](https://chickpea.co/.well-known/agent-skills/index.json).
- [Chickpea CLI](https://github.com/pejmanjohn/chickpea/tree/main/packages/cli#readme): the official
  command line, `chickpea-cli` on npm, Node 22.19 or newer. `npx chickpea-cli doctor https://<deployment>`
  checks the OAuth and MCP surface with no sign-in; `mcp config` prints the client snippet for Claude Code,
  Codex, or Cursor; `login` runs the same OAuth 2.1 public-client flow and stores tokens locally (0600);
  `workspace inspect`, `tools list`, `call <tool>`, and `recipe export|preview` reach every management
  tool from a script. It never accepts a token as an argument and never confirms a proposal by itself.
- [Comparisons index for agents](https://chickpea.co/compare/llms.txt): the section-level llms.txt.
- [Chickpea authentication design](https://github.com/pejmanjohn/chickpea/blob/main/docs/authentication.md)
- [Chickpea documentation](https://docs.chickpea.co)
- [Chickpea configuration and environment variables](https://docs.chickpea.co/reference/environment-variables/)
- [Chickpea Slack app manifest](https://docs.chickpea.co/reference/slack-app-manifest/)
- Errors: a request to an API-shaped path on chickpea.co (/mcp, /api/*, OAuth metadata) gets a JSON
  problem (RFC 9457) with `code: not_a_deployment` and a `resolution` naming the deployment path.
- Reading this site: every page answers `Accept: text/markdown` with markdown and sets
  `Vary: Accept`; the same markdown is at the page path with `.md` appended. A wrong path returns a
  real 404 with recovery links, as markdown for a markdown client and as JSON for a JSON client.

## Comparisons

Side-by-side pages, each with a criteria table, a monthly cost estimate for a 25-person team, where
each product is better, and a FAQ. Every page has a markdown twin at the same path with `.md` appended.

- [All comparisons](https://chickpea.co/compare)
- [Chickpea vs Claude Tag (Anthropic)](https://chickpea.co/compare/claude-tag)
- [Chickpea vs OpenClaw](https://chickpea.co/compare/openclaw)
- [Chickpea vs Buzz (Block)](https://chickpea.co/compare/buzz)
- [Chickpea vs Grok Bot (xAI)](https://chickpea.co/compare/grokbot)
- [Chickpea vs Viktor](https://chickpea.co/compare/viktor)
- [Chickpea vs Hermes Agent (Nous Research)](https://chickpea.co/compare/hermes)

## Blog

Long-form posts for teams choosing or running AI teammates in Slack: alternatives, comparisons,
and how-tos. Each post is sourced and dated, re-checked every 90 days, and has a markdown twin at
its path with `.md` appended.

- [All posts](https://chickpea.co/blog)
- [OpenClaw alternatives for teams in Slack: what breaks when you share one agent](https://chickpea.co/blog/openclaw-alternative-for-teams): OpenClaw's own trust-boundary docs, the roster mechanics in Slack, four ways out including keeping OpenClaw, and the governed setup.

## About the project

- [About Chickpea](https://chickpea.co/about): built by Paperplane Labs, and how it is paid for.
- [Pricing](https://chickpea.co/pricing): free software; what a 25-person team pays its providers.
- [Contact](https://chickpea.co/contact): bugs, questions, security reports, mailing address.
- [Privacy](https://chickpea.co/privacy)

## Links

- [Site](https://chickpea.co/)
- [Source and documentation](https://github.com/pejmanjohn/chickpea)
- [Readme](https://github.com/pejmanjohn/chickpea#readme)
- [Setup guide](https://github.com/pejmanjohn/chickpea/blob/main/SETUP_AGENT.md)
- [Telemetry](https://docs.chickpea.co/security/telemetry/)
- [Runbooks](https://github.com/pejmanjohn/chickpea/tree/main/docs/runbooks)
- [Licence, Apache 2.0](https://github.com/pejmanjohn/chickpea/blob/main/LICENSE)
- [Deploy to Cloudflare](https://deploy.workers.cloudflare.com/?url=https://github.com/pejmanjohn/chickpea)
- [Flue, the agent framework underneath](https://flueframework.com)