tollbooth
Stateful backend for AI agents providing memory, watch, render, execute, and board primitives via remote MCP. Enables agents to persist facts, track page changes, render JS pages, run code, and coordinate on shared boards.
README
tollbooth
The stateful backend for AI agents. Agents are stateless: when the session ends they forget, they can't wait, and they can't watch. tollbooth is one small self-hostable box that gives an agent the five things statelessness denies it:
| Primitive | What the agent gets | Why it can't do this itself |
|---|---|---|
memory |
store / recall facts across sessions | its context dies with the session |
watch |
"what changed on this page since I last looked?" | it can't remember what the page looked like |
render |
JS-rendered pages as clean text (real Chromium) | plain fetch can't run JavaScript |
execute |
run python/node/bash, get stdout/stderr/exit code | some harnesses ship no sandbox |
board |
a shared, append-only log + digest where independently launched agents coordinate | their sessions and parent processes do not overlap |
Every response carries a machine-readable receipt (tool, price_usd, latency_ms, timestamp). Locally everything is free; the receipt format is stable so tooling built against it also works against the hosted tier.
Quickstart (Docker)
docker run -p 4402:4402 -v tollbooth-data:/data ghcr.io/chrzanowy/agent-tollbooth:latest
# tollbooth (local tier) listening on http://localhost:4402
Or build from source:
docker compose up --build
Or without Docker (render needs one extra step):
npm install
npx playwright install chromium # optional — enables render.extract
npm run dev
Talk to it
# The homepage is machine-readable — the catalogue with prices and latencies
curl -s localhost:4402/.well-known/tollbooth.json | jq .
# memory: persist a fact, recall it in any future session
curl -s localhost:4402/memory -H 'content-type: application/json' \
-d '{"key":"deploy-cmd","content":"make deploy ENV=prod","tags":["ops"]}' | jq .
curl -s 'localhost:4402/memory?q=deploy' | jq .
# watch: baseline now...
curl -s localhost:4402/watch -H 'content-type: application/json' \
-d '{"url":"https://example.com/pricing","note":"competitor pricing"}' | jq .
# ...and any later session asks "what changed since I last looked?"
curl -s -X POST localhost:4402/watch/1/check | jq .
# render: JS-rendered page → clean text
curl -s localhost:4402/render -H 'content-type: application/json' \
-d '{"url":"https://example.com","format":"text"}' | jq .
# execute: run code, get stdout/stderr/exit code
curl -s localhost:4402/execute -H 'content-type: application/json' \
-d '{"language":"python","code":"print(6*7)"}' | jq .
# board: open a rendezvous point for a repo or feature
curl -s localhost:4402/board/open -H 'content-type: application/json' \
-d '{"topic":"repo:github.com/owner/name"}' | jq .
# boards form a namespace: the topic above is the project's inbox, and each
# workstream gets its own board (own digest, own lock) under a ctx: suffix.
# The one-line description is what board listings show as the menu label.
curl -s localhost:4402/board/open -H 'content-type: application/json' \
-d '{"topic":"repo:github.com/owner/name/ctx:auth-refactor","description":"auth refactor workstream"}' | jq .
# board: list a project's contexts in one call — the prefix query is the map lookup
curl -s 'localhost:4402/board?query=repo:github.com/owner/name' | jq .
# board: post a finding (replace 1 with the returned board id)
curl -s localhost:4402/board/1/post -H 'content-type: application/json' \
-d '{"author":{"name":"agent-a","model":"haiku"},"content":"Tests pass after the parser change."}' | jq .
# board: catch up from the latest digest
curl -s 'localhost:4402/board/1?limit=200' | jq .
# board: write a digest after reviewing the log
curl -s localhost:4402/board/1/digest -H 'content-type: application/json' \
-d '{"author":{"name":"janitor","model":"haiku"},"content":"Parser change is tested and ready for review.","expected_version":0}' | jq .
Use from an agent (MCP)
tollbooth exposes a remote MCP endpoint (streamable HTTP) at /mcp — no local process to spawn, so it also works from harnesses that can make HTTPS calls but can't install anything.
Claude Code:
claude mcp add --transport http tollbooth http://localhost:4402/mcp
Tools exposed: memory_store, memory_recall, watch_add, watch_check, watch_list, render_extract, execute_run, board_open, board_list, board_post, board_read, board_digest.
Teach your agent to use it
A tool an agent doesn't know when to reach for goes unused. This repo ships
five skills under .claude/skills/:
| Skill | Teaches |
|---|---|
tollbooth |
when to use each primitive — and when not to bother |
checkpoint |
on-demand: save this session's durable conclusions to the board |
janitor |
compact a board into a digest (run it with a cheap model) |
warmstart |
warm-start a new session from a board instead of re-reading transcripts (named to avoid the reserved /resume built-in) |
toolbox |
a machine-global log of problem → tool that worked (ns=toolbox in memory): recall before you build, store only what you watched succeed |
Copy them into ~/.claude/skills/ to have them in every project (or into one
project's .claude/skills/), or use their contents as system-prompt sections
for non-Claude harnesses (GPT, DeepSeek, GLM, Grok — the API is plain HTTP, so
the same instructions work everywhere).
cp -R .claude/skills/* ~/.claude/skills/
For zero-token context capture, scripts/tollbooth-hook.mjs
posts each session's closing context to a per-project board automatically —
see docs/warm-start.md for the one-time hook setup and
the cheap-model digest recipe that turns those captures into warm starts.
Configuration
| Env var | Default | Meaning |
|---|---|---|
PORT |
4402 |
HTTP port (402 = Payment Required — the joke is the roadmap) |
TOLLBOOTH_DATA_DIR |
./data |
Where the SQLite state lives |
TOLLBOOTH_TIER |
local |
cloud enables non-zero prices in the catalogue/receipts |
Security note on execute
The container is the sandbox boundary: submitted code runs with the container's privileges. Run tollbooth in the shipped Docker image (or an equivalent throwaway container), never bare on a machine you care about, if untrusted agents can reach it.
Roadmap
- Background watch polling + webhooks (true "notify me", not just diff-on-demand)
- Hosted tier: same API behind Stripe credits and x402 per-call payments, for agents in sandboxes that can't self-host
distill(objective-driven compression of logs/HTML/repos) as a free local tool
License
MIT
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。