A coding agent
that lives where
you actually work .

Unified streaming API with cost tracking across OpenAI, Anthropic, Google, Mistral, Bedrock and more.

Read, write, edit, bash, grep, find. Everything an agent needs, with safety rails.

Branching sessions, context compaction, HTML export, full audit trail. Since 0.9.41 a session also keeps its prompt cache warm while you think: Anthropic's cache expires after five minutes, and the pause to read a diff is usually longer than that, so a tiny ping four minutes in keeps the next turn a cache read rather than a re-send of the whole conversation. Two pings, then it stops, because past a quarter of an hour you are doing something else. And a standing instruction you add mid-conversation with /system stays in the session, survives compaction, and is marked as coming from you rather than from a tool.

Custom tools, skills, prompt templates, themes and hooks — discovered automatically.

A cloned repository can no longer make Elyra run code just by being opened. Project packages, project extensions, MCP server definitions and the settings that change how commands run — shellPath, shellCommandPrefix, npmCommand, autoExtensions and autoSkills — take effect only once you trust the directory, and Elyra lists exactly what would run before it asks. AGENTS.md, skills, prompt templates, themes and every other setting work as before, trusted or not. Trust this folder, its parent, or this session only; /trust and elyra trust change your mind later. Print, JSON, RPC and MCP modes never prompt, so in a script an untrusted project simply runs without its gated configuration unless you pass --trust-project.

Elyra uses the MCP servers you already configured for Claude Code, Cursor or VS Code, through two stable tools, mcp_search and mcp_call, instead of loading every server's schemas into the prompt; with no servers it adds nothing and costs nothing, and it comes with the npm install (the standalone binary builds run elyra install npm:@elyracode/mcp once). An app that starts Elyra can hand it servers for that one process in the ELYRA_MCP_SERVERS environment variable, in the .mcp.json format (Elyra Workspace passes its agent gateway this way), and it is read in every project, trusted or not, since whoever sets it already controls the process. The other way, elyra --mode mcp serves Elyra to Claude Desktop, Cursor and other hosts: elyra_task runs the agent read-only unless you say edit, elyra_review reviews your uncommitted changes with a model from another vendor, and elyra_decisions lists the recorded architecture decisions.

Differential rendering, markdown, syntax highlighting, image and editor support.

Lit components for embedding chat, artifacts and model pickers in your app.

Auto-selects cheap models for simple tasks, powerful ones for complex work.

If a provider is down or rate-limited, turns auto-retry on an equivalent model from another vendor. /review --cross gets a second opinion from a different vendor entirely.

Elyra records per-model outcomes — turn success, edit failures, retries, failovers, latency, cost — to a local ledger and demotes demonstrably unreliable models in failover ranking. Since 0.9.37 the ledger also recommends: /routing-stats ends with one concrete suggestion per tier — switch a pinned model when a measured alternative is meaningfully better, which means a higher success rate or equal success at under 60% of the cost per turn. And every commit the agent makes carries an Elyra-Cost: trailer — spend since the previous commit, the model, the turns — so a team reads AI cost per change in git log (gitCostTrailer: false to stop). Since 0.9.38 routing picks the thinking level per turn as well as the model; your session level is a ceiling it only ever lowers. Only ids, counters and timings; never prompts or code, nothing leaves disk. Inspect with elyra doctor routing.

Persists project knowledge across sessions — no re-explaining your stack.

TALL, VILT, RILT and more. Auto-detected. Drop-in expert context per stack.

/init detects your stack, creates AGENTS.md, blueprints and snippets directories, and suggests extensions.

/snippet sends saved instructions from .elyra/snippets/ — release flows, review checklists, canned prompts.

/swarm build, review and refactor pipelines chain focused agent stages with visual progress.