An agent that lives in your terminal and does the work — not a chat window that describes it.
What it does
Solves a task end to end
It plans, edits files, runs the verification command, and reverts the whole attempt if the check fails. A run that changed nothing on disk is reported as a failure, not a success — and every receipt names which authority approved it: a test, a diff, a reviewing model, or nothing.
Splits work and isolates it
Several tasks run at once, each in its own git worktree, so parallel agents cannot edit the same file behind each other's backs. Safe edits are merged; conflicts are reported rather than guessed at.
Remembers, and schedules
Short-term, recent, factual and about-you memory with a relation graph and a full-text store, plus cron-expression scheduling — the job itself written in plain language — so it runs without being asked.
Reaches the world
Files, a shell, a real rendered browser, any MCP server, web search, image generation, speech, media download, data analysis and charts. Messaging on Discord, Telegram, Slack, Signal and WhatsApp, plus an HTTP endpoint.
Learns, and keeps only what passed
When it succeeds at the same kind of task more than once, that becomes a reusable skill card. A card born from a run that consumed untrusted content is held pending, out of retrieval, until someone approves it with chimera skills-approve.
Install it
One command and one key for the core. Reading documents, transcribing audio, drawing charts, downloading media and talking to MCP servers come with the [full] extra; web search, image generation and voice each want their own key. chimera features is the live checklist.
pip install chimera-agentpip install 'chimera-agent[full]'chimera initchimera solve "add a --json flag to the report command" --verify "pytest -q"The fence is not on by default. --guard puts every tool call through the governance kernel and --taint marks what came from outside; neither is the default. The default sandbox is local and does not isolate — for autonomous work, or code you did not write, set CHIMERA_SANDBOX=docker. With no interactive terminal (cron, CI), a host command is refused by default, and the scheduler is still one of the surfaces that does not track taint.
Bring your own model
One OpenRouter key reaches a hundred models. Native keys for OpenAI, Anthropic, Gemini and DeepSeek work directly, several keys per provider rotate on rate limits, and a local model through Ollama needs no key at all. Any OpenAI-compatible endpoint works by pointing one setting at it.
export CHIMERA_MODEL=ollama/llama3.1