Introducing CoreSpeed TIE
Lend your agents your computer. Any agent, anywhere, works the way you do.
Blog
Notes from building the harness AI agents run on — connectors, memory, budgets, billing, and the engineering in between.
RSS feedProduct
Launches and what shipped.
The media tool now speaks
Text to speech, voice design and transcription from Fish Audio are now part of CoreSpeed's media tool, on the same MCP endpoint your agents already use.
TinyFish, now on CoreSpeed
Web agents, search, saved sign-ins and monitors from TinyFish, in every agent you use, through one CoreSpeed connection.
CoreSpeed for SEO
Keyword, rank and audit data for any agent, billed per call in credits.
Xbrain: Give Your Agent a Sense of Your Taste
Turn your X bookmarks into searchable design context for your agent.
Introducing Smart Approval
Decide which of your agents' actions need your sign-off. Everything else keeps running; the ones you named wait for you in Slack, by email or on the dashboard.
Introducing CoreSpeed
An operating system for agents.
Concepts
What the words mean, and why they matter for agents.
What are spend caps for AI agents, and how do they work?
Spend caps stop an agent loop from draining a budget. On CoreSpeed an org wallet threshold and per-key monthly caps refuse metered calls before a tool runs.
What does human-in-the-loop mean for AI agents in 2026?
Human-in-the-loop means a person decides named actions before they run, under a policy written in advance. Reads never wait, and an approved call runs once.
What is a plain-English approval policy?
Sentences, read in full for every write an agent makes. Every save is a version, Test policy runs a hypothetical call, and a default you pick covers the gaps.
What is a remote MCP server, and how do you connect one?
A remote MCP server is reached over HTTPS instead of run as a local process. How sign-in works, and how to add one in Claude Code, Codex and Cursor.
What is a shared connection, and when should you share one?
A shared connection is an app account every member and agent in the org can act through. Private is the default. Here is who can share, remove and reach one.
What is agent memory? Shared, private, and owned by the org
Agent memory is durable, searchable state that outlives one conversation, client or agent. How it differs from a context window, and who owns it.
What is an agent identity (and why not just an API key)?
An agent identity is a principal of its own, with an id, keys, an owner who answers for it, and reach limited to shared connections. A personal API key acts as you.
What is an audit trail for AI agents?
An audit trail records what each agent did, as whom, and how it ended. CoreSpeed keeps one per org: action, actor, outcome, code and cost, never the prompt.
What is an MCP gateway, and why do agents need one?
An MCP gateway is one remote MCP server that fronts many tools and app accounts behind one sign-in. What that changes for config, credentials, spend and audit.
What is credential custody for AI agents?
Credential custody means the agent asks for an action and the platform performs it with the right account. The agent never holds the token, so a leak exposes no app credential.
What is Harness-as-a-Service (HaaS)?
Harness-as-a-Service is a hosted harness any agent signs into. Tools, auth, memory, budgets and an audit trail come as one service, and the agent you run stays yours.
What is MCP, explained for teams running several agents
MCP is the JSON-RPC protocol an agent uses to list and call tools on a server. For teams, one remote server means shared connections, one policy and one ledger.
What is SKILL.md, and why is setup one line?
SKILL.md is the root file of an agent skill: a name, a description and instructions. CoreSpeed publishes its setup as one, so a single chat line installs it.
Why tools/list is yours, and not a catalog
tools/list on CoreSpeed returns exactly what this sign-in can call. Absence is a decision, presence is neither health nor permission, so never hard-code it.
Guides
Set up a client, connect an app, write a policy.
Connect Claude Code to Notion, Slack and GitHub with one MCP server
Add the CoreSpeed MCP server to Claude Code at user scope, sign in once, connect Notion, Slack and GitHub in the dashboard, and verify with manage__whoami.
How to add one MCP server to Cursor for apps, memory and web search
Add CoreSpeed to ~/.cursor/mcp.json, click Needs login, and Cursor gets your connected apps, durable memory, live web search and media tools from one server.
How to connect ChatGPT to your apps with a custom connector
Turn on Developer mode in ChatGPT (Plus and Pro on the web), create a connector for api.corespeed.io/mcp with OAuth, and sign in to reach your apps and memory.
How to connect Codex to your apps and give it memory
Register CoreSpeed in Codex with codex mcp add, sign in, connect your apps, and add memory rules to AGENTS.md so Codex recalls decisions across sessions.
How to connect two accounts of the same app to one agent
Connect two Slack workspaces or two X handles under one connector, give each an alias, and pass the account argument so each call lands in the right one.
How to give your agent memory that survives sessions and clients
Two standing rules, a scope choice, and a habit of saving decisions give any agent memory that outlives the session and the client it ran in.
How to register your own MCP server for your whole organization
An org admin registers any HTTPS MCP server as a remote connector; its tools join tools/list under the org__ prefix with the same sign-in, caps and trail.
How to run a headless agent or CI job with a CoreSpeed API key
Choose a member key (sk-cs-) or an agent key (sk-csa-), read it from CORESPEED_API_KEY, cap its monthly spend, and verify the key with memory__list.
How to set a monthly budget for an agent
Give the agent its own key with a monthly spend cap in credits, know what the refusal looks like, and read what each run cost in Activity.
How to share one app connection with your whole team
Connect an app once as a shared account so every teammate's agent can act through it. Who can share, who can remove, and when to keep a connection private.
How to use your apps and memory inside claude.ai and Claude Desktop
Add CoreSpeed as a custom connector on the Pro, Max, Team or Enterprise plan. It follows your Claude account to desktop and mobile with your apps and memory.
How to write your first approval policy (with examples)
Write three kinds of sentence (fine, ask, never), pick a default behavior, run Test policy, connect Slack and turn the gate on. Reads are never gated.
The cs CLI in ten minutes: login, keys, tools, usage
Install @corespeed/cs on Node 20 or Bun, sign in with cs login, mint and rotate API keys, print MCP config, search and call tools, check usage, and read exit codes in scripts.
When to bring your own OAuth app (Google, Microsoft, Meta) and how
Google, Microsoft and Meta gate shared OAuth apps behind review. An org admin registers your own app: paste client_id and client_secret, match the scopes.
Use cases
Worked examples with real tool names.
A meeting brief from your mail, calendar and notes, in one request
How an agent builds a meeting brief from Gmail, Google Calendar and Notion through private connections, remembers your format, and never holds a token.
A support agent that issues refunds, with a person on the big ones
A worked example. A support agent reads the ticket, checks the charge in Stripe and issues the refund. Refunds over $100 pause for a person in Slack.
Competitive research: search for sources, then read what matters
A playbook for an agent that finds competitor sources with web__search, reads the few pages that matter with web__scrape, and writes the brief to Notion.
Continue a task in a different agent without re-explaining
What carries over when you move a task from Claude Code to Codex through CoreSpeed, memory, connections and policy, and what does not, the conversation itself.
Generating launch images and short videos from an agent
How an agent picks a hosted model, generates a launch image, chains it into a short video through its Drive path, and collects signed URLs. Billed only after success.
Market scans with listing tools: Zillow, Airbnb, TripAdvisor and Maps
Read Zillow, Airbnb, TripAdvisor and Google Maps as structured records. Each call collects live in 10 to 60 seconds and bills only the rows it gathered.
One org, many agents: Claude Code, Codex and a bot on one rulebook
How a team keeps each member's own client, shares app connections, runs one approval policy and one ledger, and gives bots an identity of their own.
Running an unattended job you operate with a CoreSpeed agent key
Give a cron job or server process you run its own agent identity and sk-csa key: shared connections only, a monthly cap, approvals routed to its owner.
Social listening with read-only tools and no platform accounts
How an agent tracks mentions on X, Reddit, YouTube, TikTok and Xiaohongshu through read-only tools, with no platform login, and saves findings to shared memory.
How to turn CoreSpeed into a content factory
How we made PRODUCTION INCIDENT, from Blender rehearsal and generated performances to screen compositing, sound, and the final 4K edit.
Engineering
How the pieces are built, and what we measured.
Memory that keeps its receipts
Extracted facts answered 84.4% of LongMemEval-S. Linking each fact back to the conversation it came from raised that to 89.6%, for 20% more storage.
Exactly once: how an approved agent action runs with frozen arguments
A held call is frozen, the agent waits on one id, and approval hands the call to that waiting request once. Re-sending the call makes a second card.
How memory search works: hybrid retrieval, excerpts and result budgets
memory__search fuses exact and relaxed lexical matching, entity aliases and vector similarity into one ranking, then bounds each row and the result.
HTTP 200 with isError true: the most common MCP integration bug
A CoreSpeed tool call can fail inside an HTTP 200. How the two error layers differ, which codes to retry, and which ones need a person instead.
Linking memories: a small graph that keeps decisions connected
Memory links are directed edges with a kind and optional weight. They are idempotent to write, visible only when both ends are, and capped at 1,000 per memory.
Memory isolation with row-level security, not application code
Caller identity comes from auth, each memory carries its scope, and Postgres row-level security enforces the boundary. Nothing in the request body can cross it.
Pull requests for shared memory: proposals, archive and versions
A correction next to a stale shared memory does not stop the old one being retrieved. Archive it with a successor, or propose a fix for its author to merge.
The hardest problem in multi-agent teams is scoping
Past one person, the hard question is scope: who may touch what, on whose behalf, with what budget, with what record. The organization has to answer it.
Why reads are never gated: designing an approval gate people keep on
A gate that asks about everything gets switched off. Smart Approval judges writes only, lets reads through, and never lets a failed evaluation run a write.
Why your agent should never hold an OAuth token
An agent that holds a token can leak it through a prompt, a log or a transcript. Credential custody lets a platform act with the right account instead.
How we took LongMemEval from 80.0 to 94.2 without touching retrieval
LongMemEval-S climbed from 80.0 to 94.2 while retrieval recall at 10 stayed at 0.99. The gains came from evidence budget, reader generation, and instructions.
We retrieved the memory, then dropped the answer
A 200-question conflict run scored 0.800 at the parent-Memory boundary and 0.635 at the evidence boundary. The gap was in what reached the reader.
Comparisons
One option against another, honestly.
Built-in web search vs wiring a search API into your agent
CoreSpeed's web tools return ranked sources and clean page text with no provider account, billed per request. A direct provider key gives more knobs.
Member key vs agent key vs browser sign-in: which to use
Four credentials, two kinds of principal. What each one reaches, where spend caps attach, and which credential an unattended agent should carry.
OAuth connector vs pasting an API key: which to use for each app
Most CoreSpeed connectors use OAuth. Some take a pasted provider key, a few take only that. How the index tells them apart, and how rotation and reauth differ.
One MCP server vs a config file per agent
When a per-client list of local MCP servers is enough, where it stops scaling, and what changes once a team points every client at one remote server.
Shared memory vs CLAUDE.md and AGENTS.md files
Instruction files hold the rules an agent follows, per repository or per machine and per client. Org-owned memory holds searchable facts that follow you across both. Use both.
Smart Approval vs client-side permission prompts
Claude Code's auto mode and Codex's Approve for me guard one client. Smart Approval gates the tools for every client. When each one is the right choice.
FAQ
Short answers to the questions that come up.
CoreSpeed FAQ: plans, credits, clients, security, what it does not do
Short answers on what CoreSpeed is, which agents work, what is free, where tokens live, who approves, whether it runs agents, and how to leave.
needs_reauth explained: what it means and what to do
needs_reauth means an account's upstream grant can no longer be used. Tools stay visible, calls fail, and the fix is reauthorizing in the dashboard.
payment_required and key_spend_limit_exceeded: what still works
payment_required is an organization wallet hold; key_spend_limit_exceeded is one key's monthly cap. Both arrive as HTTP 200 with isError true.
Why is a tool missing from tools/list?
A tool is missing from tools/list because this caller cannot call it now: app not connected, capability off, agent key reach, or a stale session.
Company
Why we are building this.
A letter from our founders
Why we're building an operating system for agents, and what becomes possible when they have one.