Skip to content
ContextStream Agents

Specialist context agents for Claude Code, Codex and Cursor

ContextStream Agents are specialist context agents for Claude Code, Codex, Cursor and other MCP clients. Each is purpose-built and tuned to do one job very well, has to meet a quality bar in its own tests, and works from your workspace's own context: its code index, decisions, lessons and plans. Your coding agent hands one a job and gets back a short, structured result.

Each agent gets only the access its job needs, a workspace owner or admin turns it on, and it declines any task outside its job. Agents that run code do it in an isolated sandbox with no credentials inside.

15 agents in ContextStream today. Updated

What makes them context agents

Built for one job, grounded in your workspace, scoped to what the job needs.

Purpose-built

One job each, done very well

Each agent is a program built and tuned for a single job, on general-purpose models reached through ContextStream's model proxy. Its own tests set a quality bar it must meet, including declining every task outside its job: 448 test cases with known answers across the 15 agents.

Grounded

Works from your workspace's own context

An agent reads what its job needs from your ContextStream workspace: the code index, decisions, lessons, plans, docs or connected repositories, plus the context ContextStream picks for its job. It works on your project as it is, not on a guess.

Scoped and secure

Only the access its job needs

A run gets a short-lived token for one workspace, limited to the scopes its job needs, and writes nothing beyond them. A workspace owner or admin turns each agent on. Agents that run code do it in an isolated sandbox with no credentials inside.

In ContextStream today

15 agents you can turn on now.

Each page says what the agent does and declines, what it can access, how it is tested and how to call it, and shows a real run with its input, output and time.

Code

CI Fixer

failing test fix agent

CI Fixer reproduces a failing test command in a public repository and returns the smallest change to the project's code that makes it pass, with the passing output.

Checked on 13 test cases with known answers: all 13 passv0.1.0
Code

Code Oracle

code search agent

Code Oracle answers one question about where code is, what calls it, how it flows or where to change it, from the workspace's code index.

Tested on 212 real code questions, replaying recorded live runs: 202 passv0.2.0
Data

Data Analyst

CSV and Excel analysis agent

Data Analyst answers questions about one CSV or Excel table, in plain English or structured, with charts, and recomputes every number a second way before it reports it.

Checked on 9 test cases with known answers: all 9 passv1.1.0
Code

Dependency Upgrader

dependency upgrade agent

Dependency Upgrader upgrades a public GitHub repository within its semver ranges, fixes the project's code when that breaks its tests, and returns the diff and the results.

Checked on 10 test cases with known answers: all 10 passv1.0.0
Code

Docs Drift

documentation drift agent

Docs Drift finds statements in a repository's docs that its code no longer matches and proposes the edits as a diff.

Checked on 7 test cases with known answers: all 7 passv0.1.0
Workspace

Lens

screenshot and image explainer agent

Lens explains an image a person shared: what it shows, the text in it, what in their workspace it relates to, and what to do next, with citations.

Checked on 6 test cases with known answers: all 6 passv0.2.0
Mail and briefings

Mail

inbox triage and reply drafting agent

Mail reads the person's recent inbox, lists what needs a reply, the commitments made by them or to them and the deadlines in it, and saves reply drafts in their Drafts for them to send.

Checked on 13 test cases with known answers: all 13 passv0.1.0
Workspace

Memory Gardener

workspace cleanup agent

The Memory Gardener finds duplicate or conflicting decisions and lessons and stale open items, and proposes merging, superseding or closing them.

Checked on 12 test cases with known answers: all 12 passv0.1.0
Documents

PDF

docs to PDF agent

PDF lays out a source the person names (a doc, a plan, notes or markdown) as a PDF and stores it in the workspace.

Checked on 11 test cases with known answers: all 11 passv0.1.0
Workspace

Plans & Tasks Keeper

plans and tasks status agent

Plans & Tasks Keeper answers one question about the status of the workspace's plans and tasks, citing their ids.

Tested on 106 real-world questions with the live model: all 106 passed: right plan, every cited fact checked against the data, and out-of-scope questions declinedv0.2.0
Code

PR Reviewer

pull request review agent

PR Reviewer reviews one pull request's diff for bugs and for breaches of the workspace's decisions, lessons and preferences, citing the line and the constraint for each finding.

Checked on 13 test cases with known answers: all 13 passv0.1.0
Code

Release Notes

release notes agent

Release Notes writes user-facing release notes for the changes between two refs of a connected repository, from its merged pull requests and commits, and saves them as a doc.

Checked on 13 test cases with known answers: all 13 passv0.1.0
Code

Repo Guide

repository onboarding guide agent

Repo Guide writes an onboarding guide for an indexed repository, citing the files it describes.

Checked on 7 test cases with known answers: all 7 passv0.1.0
Mail and briefings

Research Brief

feed research brief agent

Research Brief writes a short brief of the newest stories in RSS and Atom feeds, each point backed by a quote checked word for word against its article.

Checked on 7 test cases with known answers: all 7 passv1.0.0
Code

Security Audit

dependency vulnerability audit agent

Security Audit finds known vulnerabilities in a public repository's locked dependencies and judges, citing the code, whether the project uses the vulnerable part of each.

Checked on 9 test cases with known answers: all 9 passv0.1.0
How it works

Your coding agent hands off the job. The specialist does only that.

  1. Turn it onAn owner or admin turns each agent on in the workspace's agent library.
  2. Ask from your agentClaude Code, Codex or Cursor calls it through ContextStream's MCP, or runs csa run in its shell.
  3. Get a short resultThe agent works from your workspace's context, with only its own scopes, and returns a short, structured result.
Connect Claude Code, Codex, Cursor and other MCP clients:
curl -fsSL https://contextstream.io/scripts/mcp.sh | bash
Or install csa, with ContextStream's command line:
curl -fsSL https://contextstream.io/cs/install.sh | sh
Questions

About the agents.

What are ContextStream Agents?

ContextStream Agents are specialist context agents for Claude Code, Codex, Cursor and other MCP clients. Each is purpose-built and tuned to do one job very well, has to meet a quality bar in its own tests, and works from your workspace's own context: its code index, decisions, lessons and plans. Your coding agent hands one a job and gets back a short, structured result. Each agent gets only the access its job needs, a workspace owner or admin turns it on, and it declines any task outside its job. Agents that run code do it in an isolated sandbox with no credentials inside.

How are they different from a general coding agent?

A general coding agent tries whatever you ask. A ContextStream agent does one job: it is built and tuned for that job, tested against a quality bar for it, grounded in your workspace's context, and limited to the access that job needs. It declines anything else. Your coding agent stays in charge and hands it the job.

Are they trained models?

No. Each agent is a program that calls general-purpose models through ContextStream's model proxy, with prompts tuned against its own tests. Your workspace's context is read for the run that needs it and is not used to train models.

How are the agents tested?

Each agent has its own test suite of cases with known right answers: realistic jobs, inputs it must refuse with a clear error, and requests outside its job it must decline. A case passes only if the run ends the way the case expects and every check on it holds, such as every fact exact, every quote found in its source or every cited line real. Most suites run on fixed test data with a stand-in for the model, which checks everything ContextStream builds around the model rather than the live model's own answers; Code Oracle's replays recorded runs of the live model. Each agent's page says what its cases are and what they check.

Is it safe to turn an agent on?

Each agent's page lists exactly what it can access. A run gets a short-lived token for one workspace, limited to the scopes its job needs; it works in one project, only for that run, and declines any task outside its job. 7 of the 15 agents are read-only, and the rest can write only what their scopes name, such as one note or one doc per run. Agents that run code do it in an isolated sandbox with no credentials inside. A workspace owner or admin turns each one on.

Which agents work with Claude Code and Codex?

All 15. ContextStream's MCP gives Claude Code, Codex, Cursor and other MCP clients an agents tool that finds an agent, runs it and follows the run, and csa run <agent> does the same from a terminal or an agent's shell.

Is there a code search agent?

Yes: Code Oracle. Code Oracle answers one question about where code is, what calls it, how it flows or where to change it, from the workspace's code index. More about Code Oracle

Is there a pull request review agent?

Yes: PR Reviewer. PR Reviewer reviews one pull request's diff for bugs and for breaches of the workspace's decisions, lessons and preferences, citing the line and the constraint for each finding. More about PR Reviewer

Is there a release notes agent?

Yes: Release Notes. Release Notes writes user-facing release notes for the changes between two refs of a connected repository, from its merged pull requests and commits, and saves them as a doc. More about Release Notes

What do the agents cost?

Each agent has a price per run in ContextStream credits. A priced run starts only once its price is accepted, and is charged only when it succeeds; a declined, failed, cancelled or timed-out run costs nothing. Code Oracle and Plans & Tasks Keeper are free. The others cost 400 to 8,000 credits a run. Credits come with every plan; Starter is free with 10,000 a month.

How do I turn an agent on?

An owner or admin of your ContextStream workspace turns each agent on in the web app's agent library. Agents are opening workspace by workspace; if yours is not open yet, the library says so.