llms.txt
+22 −3
https://contextstream.io/llms.txt
− - Official MCP Registry name: io.contextstream/mcp− - LongMemEval-S: 90.0% (450/500) on the full suite, official GPT-4o judge, with self-consistency; 89.6% single-shot.− - Code search: 99.38% useful-file Recall@10 on the judged corpus.+ - [ContextStream Agents: specialist context agents for Claude Code, Codex and Cursor](https://contextstream.io/agents)+ - Official MCP Registry name: io.github.contextstream/mcp-server+ - LongMemEval-S: 96.8% (484/500) on the full suite, official GPT-4o judge, Claude Opus 5.5 reader; 96.0% (239/249) on a half held out from tuning, pre-registered.+ - LoCoMo: 95.7% (1,474/1,540, categories 1-4), the EverMemOS framework's judge; 95.8% on a held-out half; 97.8% on the audited answer key.+ - Code search: 98.77% best-answer and 100% useful-file Recall@10 on the judged corpus, ContextStream auto in production on October 27, 2026 (August 1: 93.83% and 99.38%).+ ## 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+ - [CI Fixer](https://contextstream.io/agents/ci-fix): failing test fix agent. Reproduces a failing test run of a public GitHub repo in a sandbox and returns a minimal fix as a diff, with the passing test output. Never changes a test to make it pass. Checked on 13 test cases with known answers: all 13 pass.+ - [Code Oracle](https://contextstream.io/agents/oracle): code search agent. Answers where-is, what-calls, how-does-it-flow and where-to-change questions about indexed code: a cited answer of at most 600 tokens and a read list. Tested on 212 real code questions, replaying recorded live runs: 202 pass.+ - [Data Analyst](https://contextstream.io/agents/analyst): CSV and Excel analysis agent. Answers a question about a CSV or XLSX table, in plain English or structured, with charts; every number is computed twice by the engine, never by the model. Checked on 9 test cases with known answers: all 9 pass.+ - [Dependency Upgrader](https://contextstream.io/agents/dependency-upgrader): dependency upgrade agent. Upgrades a public GitHub repo within its semver ranges and runs its tests; when the upgrade breaks them, fixes the project's code. Returns the diff and the results. Checked on 10 test cases with known answers: all 10 pass.+ - [Docs Drift](https://contextstream.io/agents/docs-drift): documentation drift agent. Finds statements in a repository's README and docs that the code no longer matches, and proposes the edits as a diff, each backed by the code that shows it. Checked on 7 test cases with known answers: all 7 pass.+ - [Lens](https://contextstream.io/agents/lens): screenshot and image explainer agent. Reads an image you shared and ties it to your workspace: the text in it, related code, docs, decisions and lessons, and what to do next, with citations. Checked on 6 test cases with known answers: all 6 pass.+ - [Mail](https://contextstream.io/agents/mail): inbox triage and reply drafting agent. Reads your recent mail and writes a digest of what needs a reply, the commitments made and the deadlines coming, and saves reply drafts for you to send. It never sends anything. Checked on 13 test cases with known answers: all 13 pass.+ - [Memory Gardener](https://contextstream.io/agents/gardener): workspace cleanup agent. Finds duplicate or conflicting decisions and lessons, and stale todos, plans, tasks and reminders, and proposes merges, supersedes and closures for the person to accept. Checked on 12 test cases with known answers: all 12 pass.+ - [PDF](https://contextstream.io/agents/pdf): docs to PDF agent. Makes a PDF from a doc, a plan, notes or markdown in the workspace, with contents, bookmarks and links, and stores it as media. Checked on 11 test cases with known answers: all 11 pass.+ - [Plans & Tasks Keeper](https://contextstream.io/agents/keeper): plans and tasks status agent. Answers status, what is next, what is blocked, who is on what and what changed, from the workspace's plans and tasks, in about 300 tokens 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 declined.+ - [PR Reviewer](https://contextstream.io/agents/review): pull request review agent. Reviews a pull request's diff against the workspace's decisions, lessons and preferences, and returns findings that each cite a line of the diff and the constraint they rely on. Checked on 13 test cases with known answers: all 13 pass.+ - [Release Notes](https://contextstream.io/agents/release-notes): release notes agent. Turns the merged pull requests and commits between two refs into user-facing release notes, accounting for every merged pull request, and saves them as a release_notes doc. Checked on 13 test cases with known answers: all 13 pass.+ - [Repo Guide](https://contextstream.io/agents/guide): repository onboarding guide agent. Writes an onboarding guide for an indexed repository: architecture, where things live, how to build, run and test it, with file:line citations and a diagram. Checked on 7 test cases with known answers: all 7 pass.+ - [Research Brief](https://contextstream.io/agents/research-brief): feed research brief agent. Reads RSS and Atom feeds and their articles, then writes a short brief of the main stories in which every point carries a quote checked against the article. Checked on 7 test cases with known answers: all 7 pass.+ - [Security Audit](https://contextstream.io/agents/audit): dependency vulnerability audit agent. Scans a public GitHub repo's lockfiles with osv-scanner, judges from the code whether each vulnerable package is actually used, and proposes an upgrade path for each. Checked on 9 test cases with known answers: all 9 pass.