HarnessMatch
Menu

Extensible harness

Pi Agent

Minimal terminal harness designed to be extended, embedded, or automated.

Status
Active
License
MIT
Evidence
Documented + code-verifiable, 19 sources
Product record checked
2026-07-27
Compare harnessesInspect sources

At a glance

An MIT-licensed terminal coding agent with a small core, broad provider and local-model support, JSON and RPC modes, and a TypeScript extension system for users who prefer to assemble their own workflow.

Good choice if

  • Developers who want a small terminal agent instead of a bundled platform
  • Custom workflows built with TypeScript extensions, RPC, or JSON events
  • Switching among subscriptions, API providers, and local endpoints

Check before choosing

  • There is no built-in permission system or OS sandbox: filesystem, process, and network tools run with the launching user's privileges, while project trust only gates project extensions and is not an execution boundary
  • Subagents, MCP, browser automation, file checkpoints, plan mode, and sandboxing are extension or external-container patterns rather than core capabilities
  • Session JSONL, resume, tree branching, and compaction recover conversation state but do not roll back files, processes, or external side effects
See 2 more considerations
  • The minimal core shifts safety, orchestration, and integration work to the user; third-party extensions execute code and must be audited separately
  • Project-owned behavioral eval assets are useful for development but are not an independent product benchmark and no score is imported

Why it qualifies as a coding harness

This confirms category fit, not product quality. Every required criterion links back to first-party evidence.

Qualifies4 of 4 required criteria evidenced

  • Adaptive agent loop

    Documented

    The system repeatedly observes results and chooses the next action instead of following a fixed one-pass graph.

    EvidenceQuickstart

  • Repository tool execution

    Documented

    The system can use tools to inspect and change a repository or its execution environment.

    EvidenceQuickstart

  • Task-aware context management

    Documented

    The runtime assembles, updates, compacts, retrieves, or persists task-relevant context while work proceeds.

    EvidenceCompaction

  • Model-independent runtime control

    Documented

    Permissions, budgets, interruption, policy, or stop controls operate outside the model's own text generation.

    EvidenceRPC mode

Membership establishes category fit only. It does not score quality, safety, autonomy, model capability, or benchmark performance. · Read the membership rule.

How it works under the hood

Seven mechanisms mapped from first-party records. These labels describe what the harness provides, not how intelligent its model is.

7/7layers documented
  • Execution & isolationHost processDocumented mechanism, not a performance score.
  • Tooling & integrationsBuilt-in toolsDocumented mechanism, not a performance score.
  • Context & stateManaged contextDocumented mechanism, not a performance score.
  • Lifecycle & recoverySession resumeDocumented mechanism, not a performance score.
  • ObservabilitySession onlyDocumented mechanism, not a performance score.
  • VerificationManualDocumented mechanism, not a performance score.
  • Governance & permissionsHost accessDocumented mechanism, not a performance score.

Public code audit

5/5public artifacts present
Security policy
Present at inspected commit
CI workflow
Present at inspected commit
Automated tests
Present at inspected commit
Evaluation assets
Present at inspected commit
Contributor documentation
Present at inspected commit

The current tree contains 394 test-like files, ten workflows, and a project-owned behavioral eval package. Its live-provider tasks are valuable development assets but do not supply an independent, immutable model × harness × environment × budget × attempts result, so no product score is imported.

Inspect commit cee5ff7520d8828bed9955ef00419e995d1f91e0, checked 2026-07-27

Measured configurations

No benchmark run passes the full metadata admission policy for this harness yet. Missing data is not scored as zero.

Benchmark policy and all runs

Capability support

Documented first-class product support, checked against the sources below.

External tools (MCP)
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Local models
DocumentedLocal or self-hosted model pathSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Agent parallelism
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Runs without an open UI
DocumentedNon-interactive or automation surfaceSource · checked 2026-07-27The source establishes the mechanism, not its quality or availability in every mode.
Browser control
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.
Isolated execution
No built-in supportHost execution; external containerization is user-suppliedSource · checked 2026-07-27The official security guide documents no built-in permission prompts or sandbox.
Undo file changes
Not documentedNo first-class support established by the current recordAbsence of current documentation is not proof that the capability is impossible.

Primary evidence

Each capability claim is tied to a first-party record and a verification date.

Product record checked 2026-07-27
Pi Agent Harness repositoryPinned 0.82.1 source, minimal core, package layout, test assets, license, and security postureOfficial repositoryPi 0.82.1 releaseCurrent stable package release and distribution version verificationOfficial announcementQuickstartInstallation, terminal workflow, provider connection, built-in tools, and basic session useOfficial docsCommand-line usageInteractive, print, JSON, RPC, session, model, tool, and resource-loading flagsOfficial docsProvidersSubscription logins, API providers, custom endpoints, and local-model runtimesOfficial docsLocal llama.cpp modelsBundled local-model router, model download and lifecycle, context, and device configurationOfficial docsCustom providersOpenAI-compatible and custom endpoint configuration, authentication, headers, and model metadataOfficial docsSecurityHost execution, absence of permission prompts and sandbox, project trust, and external isolation patternsOfficial docs
View 11 additional sources
Additional first-party evidence11 sources
ContainerizationWhole-process Docker isolation pattern and distinction from built-in harness controlsOfficial docsSessionsAutomatic JSONL persistence, resume, branching, forks, cloning, export, and sharingOfficial docsSession formatTree-structured JSONL entries, parent linkage, tool results, metadata, and versioningOfficial docsCompactionAutomatic and manual context compaction, summaries, branch handling, and information lossOfficial docsJSON event streamMachine-readable headless events for integrations and custom interfacesOfficial docsRPC modeBidirectional JSONL control, prompts, tools, session statistics, and host integrationOfficial docsSDKEmbeddable agent sessions, model runtime, session management, tools, and event subscriptionsOfficial docsExtensionsTypeScript hooks, tools, commands, UI, custom workflows, trust, and code-execution scopeOfficial docsSkillsProject and user skills, discovery, validation, and instruction loadingOfficial docsPackage ecosystemInstallable extension bundles, local packages, distribution, trust, and update behaviorOfficial docsPi evaluation packageProject-owned behavioral evaluation harness, provider selection, and non-independent scopeOfficial repository