Claude in production, with tools and data access you can audit.
We build applications and agents on the Claude API, write MCP servers that expose your systems safely, and plan model upgrades as Anthropic ships and retires versions.
Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 (22 Sep to 7 Oct 2026)
Next retirement
Claude Sonnet 4.5 leaves the Claude API on 30 Nov 2026
Out of beta
Files API and Skills API, 19 Aug 2026
release watch
Claude 5.5 models07 Oct 2026
Claude Sonnet 4.5 deprecation30 Sep 2026
connects with
in short·15-second read
Cohash builds production applications and agents on Anthropic's Claude API.
We design tool use and MCP servers for your internal systems, add evaluation and logging, and scope credentials so agents act only as permitted. We also handle model upgrades. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 launched between September and October 2026, and Claude Sonnet 4.5 retires from the API on 30 November 2026.
you probably need this if
Your application calls a Claude model that has a published retirement date.
Agents need to reach internal systems, and so far the only option anyone's offered is a shared admin key.
A few teams are writing MCP servers, and every one of them does auth and logging its own way.
19 Aug 2026Files API and Agent Skills out of beta You can drop the beta headers. The Files API and the Skills API work on the Claude API without them now. source ↗
30 Sep 2026Claude Sonnet 4.5 deprecation Announced 30 September 2026. Retirement from the Claude API is scheduled for 30 November 2026. source ↗
07 Oct 2026Claude 5.5 models Claude Opus 5.5 launched 22 September, Claude Sonnet 5.5 on 28 September and Claude Haiku 5.5 on 7 October 2026. source ↗
today · Oct 22, 2026
no date yetMCP specification 2026-07-28Moves the Model Context Protocol to stateless, self-contained requests, with Tasks and MCP Apps as opt-in extensions. MCP is governed by the Agentic AI Foundation under the Linux Foundation.source ↗
the story · anthropic claude estate
From the first call to a setup your own team can run.
Here's what usually prompts a call about your Anthropic Claude estate, what we do about it and in which order, and where you end up.
01 · where it hurts02 · what we change03 · how it runs04 · where you end up
incident board · anthropic claude estate3 open
openYour application calls a Claude model that has a published retirement date.
recurringAgents need to reach internal systems, and so far the only option anyone's offered is a shared admin key.
blockedA few teams are writing MCP servers, and every one of them does auth and logging its own way.
exposure
Model upgrades and retirementsMCP server developmentAgents and tool useRetrieval and document workflowsEvaluation, cost and monitoring
startrehearsehandover
Inventory models and tools
Build the evaluation set
MCP servers behind OAuth
Agents with limits
Upgrade and hand over
we are here
ready to hand over
All services run on a current model.
Ask security which agent called which tool for which user, and they can answer.
Inventory of every Claude model ID in your code, with its deprecation date
Evaluation set built from real tasks, with graders and baseline scores per model
chapter 01 · where it hurts
Why teams reach out
Your application calls a Claude model that has a published retirement date.
Agents need to reach internal systems, and so far the only option anyone's offered is a shared admin key.
A few teams are writing MCP servers, and every one of them does auth and logging its own way.
chapter 02 · what we change
How we solve it
Model upgrades and retirements
MCP server development
Agents and tool use
Retrieval and document workflows
Evaluation, cost and monitoring
chapter 03 · how it runs
Our approach
Inventory models and tools
Build the evaluation set
MCP servers behind OAuth
Agents with limits
Upgrade and hand over
chapter 04 · where you end up
What changes
All services run on a current model.
Ask security which agent called which tool for which user, and they can answer.
Inventory of every Claude model ID in your code, with its deprecation date
Evaluation set built from real tasks, with graders and baseline scores per model
what we do
Our Anthropic Claude work
5 kinds of Anthropic Claude work, each run by engineers who have done it on live tenants.
Applications moved to current Claude models before retirement dates, with breaking API changes handled and results compared on your own evaluation set.
MCP servers that expose your internal APIs and data as tools, with OAuth-based authorization, per-tool permissions and logging of every call.
Agents built on the Messages API, Agent Skills or Managed Agents, with clear limits on what each tool can change and human approval where it matters.
Claude working over approved documents through the Files API or your own retrieval, with citations so reviewers can check every answer.
Test sets from real tasks, prompt caching and model choice tuned for cost, and production telemetry for latency, errors and spend.
capability 01 / 05in production
Model upgrades and retirements
Applications moved to current Claude models before retirement dates, with breaking API changes handled and results compared on your own evaluation set.
We pieced these together from Anthropic Claude situations that come up a lot. They're patterns. You won't find a specific client in them.
case file 01
Sonnet 4.5 pinned in six services, weeks before its retirement date
composite · not a client story
Beforethe situation
Six internal services call claude-sonnet-4-5-20250929 directly, two with thinking settings tuned by hand. Nobody kept test cases for the prompts, and the 30 November 2026 retirement is already on the calendar.
What we dothe work
First we search every repo and gateway log for model IDs and give each one an owner. Then we take a sample of last month's production requests and turn it into a graded test set. Each service runs against Sonnet 5.5 behind a flag, and we fix the prompts whose scores drop.
Afterwhere you end up
All services run on a current model. Each one has a score from its own test set, so there's something to compare against next time, and when the next deprecation notice arrives it goes into the sprint as an ordinary ticket.
case file 02
Five teams, five MCP servers and one shared service account
composite · not a client story
Beforethe situation
Product teams wrote MCP servers for Jira, the warehouse and an internal pricing API. Every one of them hits the backend with the same long-lived service account, and none of them records which user or agent made a tool call.
What we dothe work
The servers go behind your Okta or Entra ID authorization server. Tokens are issued with per-tool scopes and the user's identity is passed through to the backend. Then we add structured call logging and move clients that support the 2026-07-28 spec onto it.
Afterwhere you end up
Ask security which agent called which tool for which user, and they can answer. One compromised server can't reach past its own scopes. The next team to build a server starts from the shared template.
We don't start a step until the last one has produced something you can check.
step 01
Inventory models and tools
We go through your applications and write down every model ID, tool and MCP server they use. Then we check those against Anthropic's deprecation dates.
step 02
Build the evaluation set
Real tasks become a graded test set. That way a new Claude model or prompt gets accepted on results, rather than on how it seemed.
step 03
MCP servers behind OAuth
We expose internal systems through MCP servers. Each one checks the OAuth token from Okta or Entra ID, grants permissions tool by tool and logs every call.
step 04
Agents with limits
Give every agent its own identity. Hand it only the tools it needs, and put a human approval step in front of changes that matter.
step 05
Upgrade and hand over
We run each model upgrade through the evaluation set before it goes to production, and leave your team the runbook for the next retirement notice.
ai on anthropic claude
Where AI fits on Anthropic Claude, and where it needs guardrails.
Anthropic makes the Claude models and the API parts you build agents with. Tool use runs through the Messages API. Agent Skills and the Files API are out of beta, there's Managed Agents, and the MCP connector is still in beta. Opus 5.5, Sonnet 5.5 and Haiku 5.5 shipped in September and October 2026. Most of the engineering work sits in MCP server security and in re-running evaluations when models change.
AI agentasks for data or an action
Scoped identityonly the access it needs
MCP serverthe one door into the platform
Logged, approveda person signs off on writes
Anthropic Claudedoes the work
Anthropic created MCP, and Claude is the client. It gets to your systems through MCP servers you run, either called from your app or through the MCP connector. Each of those servers should check OAuth tokens carrying the user's identity and hand out scopes per tool. Log every call. Anything that changes data waits for a human to approve it.
Claude 5.5 models and the Sonnet 4.5 retirement
Opus 5.5 shipped on 22 September, Sonnet 5.5 on 28 September and Haiku 5.5 on 7 October 2026. Sonnet 4.5 leaves the API on 30 November 2026. We find every pinned model ID, run your evaluation set on the replacement and fix any prompt whose score drops.
MCP connector in the Messages API
The MCP connector is still beta, behind the mcp-client-2025-11-20 header. With it the Messages API calls remote MCP servers directly, using tool allow and deny lists and OAuth bearer tokens. It isn't covered by zero data retention. We check that against your data rules before anyone turns it on.
Agent Skills and Files API
Skills API and Files API are out of beta, so the headers can go. We write your repeatable procedures up as skills, keep them versioned in your repo, and review what files and scripts each skill can reach.
MCP servers you control
Two changes in MCP: the 2026-07-28 spec moves to stateless, self-contained requests, and governance now sits with the Agentic AI Foundation under the Linux Foundation. Our MCP servers check OAuth tokens from Okta or Entra ID. Every tool gets its own permission scope, and we log every call.
How Claude reaches your systems through MCP servers you control
Your application sends prompts and tool definitions to the Claude API. Tool calls go to MCP servers that require OAuth tokens from your identity provider and touch only the systems each tool is scoped to, while an evaluation set checks every model or prompt change.
01 · Application
Your app or agentMessages API · Agent SDK → Claude API · prompt + tools→ MCP servers · client-side tool calls
Evaluation setreal tasks · graders → Claude API · eval runs
These are working documents, written against your configuration, so your team can run them without us.
inventory01
Inventory of every Claude model ID in your code, with its deprecation date
yours to keep
design02
Evaluation set built from real tasks, with graders and baseline scores per model
yours to keep
code03
MCP server template with OAuth token checks, per-tool scopes and call logging
yours to keep
runbook04
Migration runbook for the next model retirement notice
yours to keep
design05
Cost report by model, feature and team, including prompt caching hit rates
yours to keep
talent and delivery · anthropic claude
Certified engineers, wherever you need them.
Some clients want a project team, some want a couple of engineers on theirs. Either works. Onsite, nearshore or offshore, with handovers across time zones that keep things moving.
Put these to anyone bidding on your Anthropic Claude work. The ones who've done it answer differently from the ones who've read about it.
Q1
Before we switch, how will you show that a newer Claude model does at least as well on our own tasks?
Model upgrades change output style, refusals and tool-call patterns. Without a graded test set drawn from your own traffic, the decision rests on a few hand-checked prompts.
Q2
Where does the user's identity go when an agent calls our internal API through MCP?
With one shared credential on the MCP server, the backend sees every tool call as the same user. Audit and least privilege only work if the token says who asked and what they're allowed to do.
Q3
Which MCP clients of ours support the 2026-07-28 specification? And what's the fallback?
The 2026-07-28 spec makes requests stateless. Mix versions between client and server and you get failures that look like network trouble, which is why the transition plan has to cover both ends.
Q4
What will you log for each tool call, and who reads those logs?
Agents change data through tool calls. If you only log the prompt you miss what the agent actually did. And a log that nobody reads won't stop anything.
Q5
How will you set a cost ceiling per feature before launch?
Token use grows with context length and agent loops. Per-feature budgets, prompt caching and a smaller model for simple steps are easier to design in than to retrofit after the first invoice.
01What do we need to do before Claude Sonnet 4.5 retires?
Find every place your code calls claude-sonnet-4-5-20250929. Move each one to a current model and run your evaluation set against it. Newer models can change API behavior, thinking settings for example, so read Anthropic's migration notes and test before 30 November 2026.
02Which MCP specification version should our servers support?
New servers should target the 2026-07-28 specification. It changes how sessions and capability negotiation work, so check which versions your clients support before you drop older ones. We often run both during a transition.
03How do you keep agents from doing too much?
We give every agent its own identity and scoped credentials, and limit its tools to the actions it actually needs. Changes that matter need approval. Every tool call goes in a log for review.
04What drives the cost of a Claude application?
Mostly model choice, tokens per request and request volume. To lower cost, use prompt caching, put smaller models such as Haiku on simple steps, and trim context. We measure real usage before suggesting changes.
05Should we build on Managed Agents or call the Messages API directly?
It depends on how much of the agent loop you want to own. The Messages API gives full control over tools, state and logging. Managed Agents hands the loop and sandbox to Anthropic, which is quicker to start but leaves less to tune. We often prototype both against the same evaluation set before choosing.
services
Related services
Projects on Anthropic Claude usually end up touching one of these, so they're worth a look.
