New issues and PRs from new contributors are auto-closed by default. Maintainers review auto-closed issues daily. See CONTRIBUTING.md.
This is the home of the Pi agent harness project including our self extensible coding agent.
- @earendil-works/pi-coding-agent: Interactive coding agent CLI
- @earendil-works/pi-agent-core: Agent runtime with tool calling and state management
- @earendil-works/pi-ai: Unified multi-provider LLM API (OpenAI, Anthropic, Google, …)
To learn more about Pi:
- Visit pi.dev, the project website with demos
- Read the documentation, but you can also ask the agent to explain itself
| Package | Description |
|---|---|
| @earendil-works/pi-telemetry | Vendor-neutral telemetry contracts, reference adapter, conformance tests, and typed schemas |
| @earendil-works/pi-ai | Unified multi-provider LLM API (OpenAI, Anthropic, Google, etc.) |
| @earendil-works/pi-agent-core | Agent runtime with tool calling and state management |
| @earendil-works/pi-coding-agent | Interactive coding agent CLI |
| @earendil-works/pi-tui | Terminal UI library with differential rendering |
For Slack/chat automation and workflows see earendil-works/pi-chat.
Pi does not include a built-in permission system for restricting filesystem, process, network, or credential access. By default, it runs with the permissions of the user and process that launched it.
If you need stronger boundaries, containerize or sandbox Pi. See packages/coding-agent/docs/containerization.md for three patterns:
- Gondolin extension: keep
piand provider auth on the host while routing built-in tools and!commands into a local Linux micro-VM. - Plain Docker: run the whole
piprocess in a local container for simple isolation. - OpenShell: run the whole
piprocess in a policy-controlled sandbox.
See CONTRIBUTING.md for contribution guidelines and AGENTS.md for project-specific rules (for both humans and agents). Longer term plans for Pi can also be found in RFCs.
npm install --ignore-scripts # Install all dependencies without running lifecycle scripts
npm run build # Refresh model data, then build all packages
npm run build:offline # Rebuild using existing model data without network access
npm run check # Lint, format, and type check
./test.sh # Run tests (skips LLM-dependent tests without API keys)
./pi-test.sh # Run pi from sources (can be run from any directory)GitHub releases include a versioned source archive covered by the release's SHA256SUMS file. Extract it and run the same build script used for the official standalone binaries:
VERSION="<release-version>"
tar -xzf "pi-${VERSION}-source.tar.gz"
cd "pi-${VERSION}"
./scripts/build-binaries.sh --offline-model-data --platform linux-x64 --out "$PWD/out"The source archive includes the generated provider model data used for the release. --offline-model-data builds with that snapshot instead of refreshing it from live provider catalogs. The script still installs dependencies, builds the monorepo, compiles the Bun executable, and stages its runtime assets. Package maintainers who provide dependencies separately can pass --skip-install --skip-deps.
We treat npm dependency changes as reviewed code changes.
- Direct external dependencies are pinned to exact versions. Internal workspace packages remain version-ranged.
.npmrcsetssave-exact=trueandmin-release-age=2to avoid same-day dependency releases during npm resolution.package-lock.jsonis the dependency ground truth. Pre-commit blocks accidental lockfile commits unlessPI_ALLOW_LOCKFILE_CHANGE=1is set.npm run checkverifies pinned direct deps, native TypeScript import compatibility, and the generated coding-agent shrinkwrap.- The published CLI package includes
packages/coding-agent/npm-shrinkwrap.json, generated from the root lockfile, to pin transitive deps for npm users. - Release smoke tests use
npm run release:localto build, pack, and create isolated npm and Bun installs outside the repo before tagging a release. - Local release installs, documented npm installs, and
pi update --selfuse--ignore-scriptswhere supported. - CI installs with
npm ci --ignore-scripts, and a scheduled GitHub workflow runsnpm audit --omit=devplusnpm audit signatures --omit=dev. - Shrinkwrap generation has an explicit allowlist for dependency lifecycle scripts; new lifecycle-script deps fail checks until reviewed.
If you use Pi or other coding agents for open source work, please share your sessions.
Public OSS session data helps improve coding agents with real-world tasks, tool use, failures, and fixes instead of toy benchmarks.
For the full explanation, see this post on X.
To publish sessions, use badlogic/pi-share-hf. Read its README.md for setup instructions. All you need is a Hugging Face account, the Hugging Face CLI, and pi-share-hf.
You can also watch this video, where I show how I publish my pi-mono sessions.
I regularly publish my own pi-mono work sessions here:
This is a fork. zkao is our long-lived working branch and is periodically rebased onto main (upstream releases). Before each rebase we snapshot zkao into a backup branch named zkao-v<version>-backup, where <version> is the upstream release zkao was synced to at that time. See AGENTS.md for the full procedure.
| Backup branch | Synced version | Date | Notes |
|---|---|---|---|
zkao-v0.75.5-backup |
v0.75.5 | 2026-05-28 | Snapshot before rebasing onto main (v0.77.0); web search support + zkao CI/release workflows |
zkao-v0.77.0-backup |
v0.77.0 | 2026-06-06 | Snapshot before rebasing onto main (v0.78.1); Gemini web-search × function-calling combination fix |
zkao-v0.78.1-backup |
v0.78.1 | 2026-06-09 | Snapshot before rebasing onto main (v0.79.0); client/provider tool-name collision fix, Codex SSE read-timeout fix, Gemini web-search tool conversion tests |
zkao-v0.79.0-backup |
v0.79.0 | 2026-06-12 | Snapshot before rebasing onto main (v0.79.1); preserves our Claude Fable 5 support commit, dropped during the rebase in favor of upstream's own Fable 5 metadata |
zkao-v0.79.1-backup |
v0.79.1 | 2026-06-20 | Snapshot before rebasing onto main (v0.79.8); dropped our cherry-picked Fable 5 adaptive-thinking test commit (superseded upstream), re-resolved web-search vs. refusal-detail conflicts |
zkao-v0.79.8-backup |
v0.79.8 | 2026-06-29 | Snapshot before rebasing onto main (v0.80.2); re-ported every fork commit onto the upstream Models-runtime refactor, which moved provider stream/convert logic from providers/*.ts into api/*.ts |
zkao-v0.80.2-backup |
v0.80.2 | 2026-07-08 | Snapshot before rebasing onto main (v0.80.3); re-resolved the agent.ts prepareNextTurn conflict; includes the streamProxy fix that relays nativeTools (native web search was silently dropped through the proxy) |
zkao-v0.80.3-backup |
v0.80.3 | 2026-07-09 | Snapshot before rebasing onto main (v0.80.6, which adds the GPT-5.6 luna/sol/terra models); dropped our signed empty-thinking implementation (upstream absorbed the same fix) and kept only its regression test |
zkao-v0.80.6-backup |
v0.80.6 | 2026-07-14 | Snapshot before rebasing onto main (v0.80.7, which adds a codex session-id clamp); preserves the new Meta (Muse Spark) provider commit alongside every prior fork commit |
zkao-v0.80.7-backup |
v0.80.7 | 2026-07-16 | Snapshot before rebasing onto main (v0.80.9); preserves the Meta Muse replayReasoning fix (skip replaying server-expiring reasoning items) alongside every prior fork commit |
zkao-v0.80.9-backup |
v0.80.9 | 2026-07-20 | Snapshot before rebasing onto main (v0.80.10); preserves the native web-search pricing fix (xAI/Meta) and the new DeepInfra provider alongside every prior fork commit |
zkao-v0.80.10-backup |
v0.80.10 | 2026-07-24 | Snapshot before rebasing onto main (v0.82.0); dropped the OpenCode Go API-widening commit absorbed upstream and re-resolved native web-search, Gemini combined-tool, Meta reasoning-replay, and DeepInfra conflicts |
zkao-v0.82.0-backup |
v0.82.0 | 2026-07-25 | Snapshot before rebasing onto main (v0.82.1); took upstream's e2e test-model retarget (gpt-5.5) over our equivalent fork hunks, and added the env-api-keys .catch fix for bundler-substituted rejecting imports (Turbopack unhandled rejections) |
zkao-v0.82.1-backup |
v0.82.1 | 2026-08-04 | Snapshot before rebasing onto main (v0.83.0); adds refusal/account-restriction classification, the model downgrade fallback runner, and the createAgentSession stream-function wrapper. Re-resolved the DeepInfra useMaxTokens conflict (union with upstream's isZai), took upstream's new "Provider stopped with: sensitive" errorMessage over our equivalent fork hunk and matched it from the classifier instead, and handled the new "pending" stop reason in terminal-event mapping |
zkao-v0.83.0-backup |
v0.83.0 | 2026-08-06 | Snapshot before rebasing onto main (v0.84.0); adds Muse Spark 1.2 as the Meta default. Re-ported every fork commit onto upstream's StreamOptions → ProviderRequestOptions split (our nativeTools and native-tool-collision helpers moved with it), took our omit-based streamProxy relay over upstream's hand-maintained allowlist (which had grown samplingParams), and unioned the servertooluse event with upstream's new "deferred" done reason. Adapted our additions to the brand-new packages/server/src/protocol.ts conformance guards: Usage.extras/cost.extras declared server-side, serverToolUse blocks dropped from the transcript projection |
zkao-v0.84.0-backup |
v0.84.0 | 2026-08-12 | Snapshot before rebasing onto main (v0.84.1); all 32 fork commits replayed cleanly with no conflicts and none absorbed upstream. The rebase target adds GPT Daybreak Blue to the Codex catalog. Note that npm run check and one OpenCode test fail identically on a pristine v0.84.1 tree once the catalogs are regenerated from live APIs: Google dropped gemini-2.0-flash, and OpenCode Zen moved grok-build-0.1 from the completions API to the Responses API |
zkao-v0.84.1-backup |
v0.84.1 | 2026-08-20 | Snapshot before rebasing onto main (v0.84.2); all 34 fork commits replayed and none absorbed upstream, but six conflicts needed re-resolution against upstream's strict tool-schema work: kept our Codex response.completed guard wrapped around upstream's now output-aware mapCodexEvents, threaded supportsStrictMode into the convertTools call feeding our Google native-search tools array (generative-ai and vertex), took upstream's getJsonSchemaToolParameters strict conversion over our equivalent openai-responses-shared hunks, unioned upstream's brand-new proxy.test.ts with our streamProxy serialization tests (add/add on the same path; our model const renamed to anthropicModel), dropped KIMI_STATIC_HEADERS (removed upstream) while keeping DEEPINFRA_BASE_URL, and unioned upstream's isDeepSeek into useMaxTokens alongside our isDeepInfra. Catalogs were regenerated from live APIs: DeepInfra now carries Qwen3.8-Max, Qwen3.8-2.4T-A95B, Qwen3.8-27B and Kimi K3 (no GLM 5.3 yet; 5.2 is the newest). Note that npm run check and one Baseten test fail identically on a pristine v0.84.2 tree once the catalogs are regenerated: Cloudflare AI Gateway renamed claude-sonnet-4-5 to claude-sonnet-4.5, and Baseten now reports GLM-5.2 as accepting image input |
MIT
